WO2025049394A1 - Aggregating disparate data representative of an adverse event for machine learning - Google Patents

Abstract

Methods, systems, and apparatus for accepting, by a training system, a plurality of sets of data elements, wherein a first set of data elements describe a first property of a first adverse event, a second set of data elements describe the first property of a second adverse event, and a third set of data elements describe the first property of a third adverse event, determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex, and in response: aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements into an aggregate set of data elements describing the first property for the first adverse event complex, and training, by the training system, a ML model.

Description

AGGREGATING DISPARATE DATA REPRESENTATIVE OF AN ADVERSE EVENT FOR MACHINE LEARNING

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Provisional Patent Application No. 63/535,280, filed on August 29, 2023, which is incorporated herein by reference in its.

TECHNICAL FIELD

[0002] This specification relates to machine learning, and more particularly to training machine learning (ML) models to predict characteristics of adverse events.

BACKGROUND

[0003] Adverse events, such as natural disasters, are increasing in both frequency and intensity. Example natural disasters can include wildfires, hurricanes, tornados, and floods, among several others. Natural disasters often result in significant loss that can include a spectrum of economic losses, property losses, and physical losses (e.g., deaths, injuries). Consequently, significant time and effort is expended not only predicting occurrences of natural disasters, but characteristics of natural disasters such as duration, severity, spread, and the like. Technologies, such as machine learning (ML), have been leveraged to generate predictions around natural disasters. However, natural disasters present a special use case for predictions using ML models, which results in technical problems that must be addressed to generate reliable and actionable predictions.

SUMMARY

[0004] This specification describes systems, methods, devices, and other techniques relating to training a machine learning (ML) model to predict characteristics of adverse events. More particularly, innovative aspects of the subject matter described in this specification relate to determining whether multiple adverse events are part of a larger, single adverse event complex, and if so, determining aggregate metrics for the adverse event complex. The aggregate metrics can be used to train an adverse event characteristic ML model. [0005] In general, innovative aspects of the subject matter described in this specification can include actions of accepting, by a training system, a plurality of sets of data elements, wherein a first set of data elements describe a first property of a first adverse event, a second set of data elements describe the first property of a second adverse event, and a third set of data elements describe the first property of a third adverse event, determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex, and in response: aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements into an aggregate set of data elements describing the first property for the first adverse event complex, and training, by the training system, a ML model that is configured to generate predictions relating to properties of adverse events using at least a subset of the aggregate set of data elements and at least a subset of the third set of data elements. Other implementations of this aspect include corresponding systems, apparatus, and computer programs, configured to perform the actions of the methods, encoded on computer storage devices.

[0006] These and other implementations can each optionally include one or more of the following features: the first adverse event and the second adverse event are one or more of temporally related and spatially related; determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex includes processing the first adverse event and the second adverse event through one or more heuristics; determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex includes processing the first adverse event and the second adverse event through an adverse event association model that generates a prediction indicating that the first adverse event and the second adverse event are associated with a first adverse event complex; actions further include determining that the third set of data elements is not associated with the first aggregate adverse event; aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements includes computing one or more of a union, a mean, a maximum, and a minimum; and actions further include obtaining, by an execution system, a fourth set of data elements that describe a second property of a third adverse event and a fifth set of data elements that describe the second property of a fourth adverse event, determining, by the execution system, that the third adverse event and the fourth adverse event are associated with a second adverse event complex, determining, by the training system and using at least a subset of the fourth set of data elements and at least a subset of the fifth set of data elements, a sixth set of data describing a second aggregate property of the second adverse event complex, and processing an input that includes at least a subset of the sixth set of data and using the machine learning model to generate a prediction relating to the second property of the second wildfire complex.

[0007] The present disclosure also provides a non-transitory computer-readable storage medium coupled to one or more processors and having instructions stored thereon which, when executed by the one or more processors, cause the one or more processors to perform operations in accordance with implementations provided herein.

[0008] It is appreciated that the methods and systems in accordance with the present disclosure can include any combination of the aspects and features described herein. That is, methods and systems in accordance with the present disclosure are not limited to the combinations of aspects and features specifically described herein, but also include any combination of the aspects and features provided.

[0009] Particular implementations of the subject matter described in this specification can be executed so as to realize one or more of the following advantages. The techniques described below can be used to produce a combination of features that can be used to more robustly train a ML model to predict behavior of an adverse event, specifically by determining related data sets representative of multiple, individual adverse events used to train characteristic ML models describing a single adverse event complex. Implementations of the present disclosure also reduce consumption of technical resources and are more time-efficient during training. For example, by combining seemingly disparate adverse events into a single adverse event, there are fewer adverse events to iterate training over. Accuracy of the resulting ML model is also improved. For example, without combining adverse events, characteristics (e.g., size, duration) would be underestimated. [0010] The details of one or more implementations of the subject matter described in this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.

[0011] FIG. 1 is a diagram of an example system for training a machine learning model (ML) to predict wildfire characteristics.

[0012] FIG. 2 is a flow diagram of an example process for training a ML model to predict wildfire characteristics.

[0013] FIG. 3 illustrates an example process for aggregating wildfire occurrences into a single wildfire complex.

[0014] Like reference numbers and designations in the various drawings indicate like elements.

DETAILED DESCRIPTION

[0015] The technology of this patent application is directed to training a machine learning (ML) model to predict characteristics of adverse events. More particularly, innovative aspects of the subject matter described in this specification relate to determining whether multiple adverse events are part of a larger, single adverse event complex, and if so, determining aggregate metrics for the adverse event complex. The aggregate metrics can be used to train an adverse event characteristic ML model. As described in further detail herein, implementations of the present disclosure process data of one or more databases of adverse events to determine which adverse events should be combined to a single adverse event complex. Implementations of the present disclosure include performing various transformations on characteristics associated with the adverse events that are to be combined to ensure that the combined characteristics are accurate.

[0016] Implementations of the present disclosure are described in further detail herein with reference to an example adverse event, which includes a wildfire. It is contemplated, however, that implementations of the present disclosure are applicable to any appropriate adverse event, such as natural disasters and extreme weather events. For example, implementations of the present disclosure can be used to train ML models and use ML models to predict characteristics of heavy rain events, flooding, hurricanes, and tornadoes, among other adverse events.

[0017] Traditionally, wildfire-likelihood and wildfire-hazard maps have shown areas where wildfires have occurred, and ML models can be trained using data relating to the wildfires. Once trained, such ML models can be used to predict various characteristics of potential future wildfires (e.g., likelihood of ignition) and/or present, on-going wildfires, such as spread rate, spread direction, size, and the like.

[0018] When training a ML model, it is important to ensure the accuracy of the training data. In some instances, training data can represent multiple wildfires. However, it can occur that multiple wildfires actually are from a single wildfire event. That is, multiple wildfires can be constituent wildfires of a larger, single wildfire event (also referred to as complex herein). In training of a ML model, if data relating to a single wildfire event is ascribed to two or more separate wildfire events, the ML model can be inaccurate. For example, rather than having one corpus of training data correctly relating to a single wildfire event, referred to as a wildfire complex herein, the training data will be split into multiple examples, potentially obscuring the relationship among the samples. Further, technical resources and time are expended in training on each of the wildfire events as separate wildfire events.

[0019] However, properly structuring the training data by determining whether data relates to multiple wildfire events or to a single wildfire complex poses technical challenges. For example, two wildfires that appear spatially disjoint can actually be the same wildfire complex (e.g., burning embers from a first wildfire ignite one or more second wildfires). Similarly, two wildfires that appear temporally disjoint can be in the same wildfire complex, but lying dormant for some time (e.g., if the wildfire spreads underground). For those reasons, among others, there is a need for technical solutions that can be used to determine whether data relates to a single wildfire or to multiple wildfires in the same wildfire complex. Once data has been properly attributed to a corresponding wildfire complex, the data can be used to train a ML model.

[0020] To provide context for the subject matter of the present disclosure, and as introduced above, ML has been leveraged to generate predictions around adverse events, such as natural disasters. For example, ML models can be used to generate predictions representative of characteristics of a natural disaster, such as likelihood of occurrence, duration, severity, spread, among other characteristics, of the natural disaster.

[0021] In some examples, one or more ML models can be trained to predict characteristics of a natural disaster using training data that is representative of characteristics of occurrences of the natural disaster, for example. The training data can include region data representative of respective regions (e.g., geographical areas), at which the natural disaster has occurred. In some examples, each ML model predicts a respective characteristic of the natural disaster. Example ML models can include, without limitation, a risk model that predicts a likelihood of occurrence of the natural disaster in a region, a spread model that predicts a rate of spread of the natural disaster in the region, a spread model that predicts a spread of the natural disaster in the region, and an intensity model that predicts an intensity of the natural disaster. Characteristics of a natural disaster can be temporal. For example, a risk of wildfire is higher in a dry season than in a rainy season. Consequently, each ML model can be temporal. That is, for example, each ML model can be trained using training data representative of regions at a particular period of time.

[0022] In further detail, the region data can include an image of the region and a set of properties of the region. More generally, the region data can be described as a set of data layers (e.g., N data layers), each data layer providing a respective type of data representative of a property of the region. In some examples, each data layer includes an array of pixels, each pixel representing a portion of the region and having data associated therewith that is representative of the portion of the region. A pixel can represent an area (e.g., square meters (m²), square kilometers (km²)) within the region. The area that a pixel represents in one data layer can be different from the area that a pixel represents in another data layer. For example, each pixel within a first data layer can represent X km² and each pixel within a second data layer can represent Y km², where X = Y.

[0023] An example, data layer can include an image layer, in which each pixel is associated with image data, such as red, green, blue (RGB) values (e.g., each ranging from 0 to 255). Another example layer can include a vegetation layer, in which, for each pixel, a normalized vegetation difference index (NVDI) value (e.g., in range of [- 1, 1]). Other example layers can include, without limitation, a temperature layer, in which a temperature value is assigned to each pixel, a humidity layer, in which a humidity value is assigned to each pixel, a wind layer, in which wind-related values (e.g., speed, direction) are assigned to each pixel, a barometric pressure layer, in which a barometric pressure value is assigned to each pixel, a precipitation layer, in which a precipitation value is assigned to each pixel, and an elevation layer, in which an elevation value is assigned to each pixel.

[0024] In general, data values for pixels of data layers can be obtained from various data sources including data sources provided by, for example, governmental entities, non-governmental entities, public institutions, and private enterprises. For example, data can be obtained from databases maintained by the National Weather Service (NWS), the United States Wildfire Service (USFS), and the California Department of Forestry and Wildfire Protection (CAL WILDFIRE), among many other entities. For example, weather-related data for a region can be obtained from a web-accessible database (e.g., through a hypertext transfer protocol (HTTP), calls to an application programming interface (API)). In another example, data stored in a relational database can be retrieved through queries to the database (e.g., structured query language (SQL) queries).

[0025] Because values across the data layers can change over time, the region data can be temporal. For example, temperature values for the region can be significantly different in summer as compared to winter.

[0026] Accordingly, the region data can include an array of pixels (e.g.,

[pi.i, ... , Pij])₅ in which each pixel is associated with a vector of N dimensions, N being the number of data layers. For example, p_t =

... ], where 1 is image data, V is vegetation data, and W is weather data.

[0027] As training data, the region data, which can be referred to as region training data, can include one or more characteristic layers that provides known characteristic data for respective characteristics of a natural disaster. The known characteristic data represents actual values of the respective characteristics as a result of the natural disaster. For example, a wildfire can occur within a region and, as a result, characteristics of intensity, spread, duration, and the like can be determined for the wildfire. Accordingly, as training data, the region data can include, for example,

are respective known (7 characteristics of a natural disaster in question.

[0028] One or more ML models are trained using the region training data. The training process can depend on a type of the ML model. In general, the ML model is iteratively trained, where, during an iteration, also referred to as epoch, one or more parameters of the ML model are adjusted, and an output (e.g., characteristic value) is generated based on the training data. For each iteration, a loss value is determined based on a loss function. The loss value represents a degree of accuracy of the output of the ML model as compared to a known value (e.g., known characteristic). The loss value can be described as a representation of a degree of difference between the output of the ML model and an expected output of the ML model (the expected output being provided from training data). In some examples, if the loss value does not meet an expected value (e.g., is not equal to zero), parameters of the ML model are adjusted in another iteration (epoch) of training. In some examples, the iterative training continues for a pre-defined number of iterations (epochs). In some examples, the iterative training continues until the loss value meets the expected value or is within a threshold range of the expected value.

[0029] To generate predictions, region data representative of a region, for which predictions are to be generated, is provided as input to a (trained) ML model, which generates a predicted characteristic for each pixel within the region data. An example output of the ML model can include pi j = [ ], where C is a characteristic predicted (P) by the ML model. Example characteristics can include, without limitation, likelihood of occurrence (e.g., risk), a rate of spread, an intensity, and a duration. In some examples, an image of the region can be displayed to visually depict the predicted characteristic across the region. For example, different values of the characteristic can be associated with respective visual cues (e.g., colors, shades of colors), and the predicted characteristic can be visually displayed as a heatmap over an image of the region.

[0030] While ML models are useful in generating predictions, natural disasters present a special use case for predictions using ML models. More particularly, various technical problems arise that must be addressed to generate reliable and actionable predictions. For example, and as discussed above, accuracy of events that training data represents is important in ensuring that the resulting ML model provide accurate predictions. In view of this, implementations of the present disclosure provide for selectively combining training data of multiple, seemingly unrelated wildfire events into a single wildfire complex.

[0031] FIG. 1 is a diagram of an example system 100 for training a ML model to predict wildfire characteristics. The system 100 can include a training system 150 and an inference system 170. The training system 150 can include a data obtaining engine 152, an association determination engine 155, a ML model training engine 160 and a model providing engine 165. The data obtaining engine 152 can obtain training data 110 that contains data examples 111-116 representative of areas in which wildfires have previously occurred.

[0032] As described herein, data examples can be organized as data layers (or simply “layers”) that include data values related to an area of interest. Each layer can be subdivided into pixels, which can be rectangular or other geometric shapes and the size of each pixel can be the same. Each pixel in a layer can contain a data value relevant to a portion of the area of interest. Each data layer can include a description of a region that has burned (e.g., 1 17a, 117b), each of which can be represented by one or more polygons, for example. Each data layer can further include timestamps (e.g., tl, . . ., t4) that indicate the time at which respective data was collected.

[0033] The data obtaining engine 152 can obtain training data examples from data sources, such as databases, using conventional data acquisition techniques such as structured query language (SQL) calls, Hypertext Transfer Protocol (HTTP) requests, call to web services through application programming interfaces (APIs), remote procedure calls, and so on.

[0034] The attribute determination engine 155 can accept one or more data examples from the data obtaining engine 152. The association determination engine 155 can use the data layers within each data example 111-116 to determine which areas shown 111-116 describe the same wildfire complex, as described further herein. The association determination engine 155 can create modified training data 120 that includes one or more aggregated training examples 121-124 that are grouped as respective wildfire complexes and that each include one or more data layers. [0035] The ML model training engine 160 can receive the modified training examples 120 from the attribute determination engine 155 and use the examples 120 to train a ML model 130. The ML model 130 can be any appropriate ML model that can be configured to accept data layers representing wildfire metrics and produce predictions related to a wildfire. For example, the ML model 130 can be a gradient boosted decision tree, a random forest, or a convolutional neural network. The ML model training engine 160 can produce a trained ML model 140.

[0036] The ML model providing engine 165 can provide the trained ML model 140 to the inference system 170. The training system can also provide descriptive information about the trained ML model 140 (e.g., weight values for nodes in a neural network) that can be used to configure an untrained ML model. The ML model providing engine 165 can make the ML model and/or descriptive information available using conventional techniques such as placing the information in a relational database, on a web server, or in a file system.

[0037] The inference system 170 can include a ML model acquisition engine 172, an inference engine 175, and a device interaction engine 180. The ML model acquisition engine 172 can accept a trained ML model 140, for example, from the training system 150. In some examples, the ML model acquisition engine 172 can accept descriptive information about the trained ML model 140 (e.g., weight values for nodes in a neural network) that can be used by the ML model acquisition engine 172 to configure an untrained ML model.

[0038] The inference engine 175 processes input through the trained ML model 140 to provide predictions of one or more wildfire characteristics during an inference time. The inference engine 175 can accept a set of wildfire-related features and produce output that results from processing the features. The features can include any wildfire-related metrics. For example, features can include features of the terrain including altitude, terrain slope, Normalized Difference Vegetation Index (ND VI), an Enhanced Vegetation Index (EV), a vegetation type descriptor and a vegetation cover metric. Features can also include weather-related metrics such as wind direction, average temperature, and humidity. The features can be organized as data layers containing pixels, where each pixel contains a data value related to sub-region of space, as described in further detail herein. [0039] The inference engine 175 can process the input using the trained ML model that is configured to generate a wildfire risk prediction output that characterizes predicted future characteristics of a wildfire, as described in further herein. The inference engine 175 can provide the predictions to the device interaction engine 180. [0040] The device interaction engine 180 can provide the predictions to devices 190. For example, the device interaction engine 180 can store the predictions in an Internet-connected storage facility such as a database, a web server or a file system. The devices 190 can use conventional data retrieval techniques to retrieve the predictions. For example, the device 190 can use SQL to retrieve the prediction from a database or HTTP to retrieve the prediction from a web server. Examples of devices 190 can include personal computers, tablets, mobile phones, servers, and the like. [0041] FIG. 2 is a flow diagram of an example process 200 for training one or more wildfire characteristics ML models. For convenience, the process 200 for training a wildfire characteristics ML model will be described as being performed by a system for training a wildfire characteristics ML model (e.g., the system for training a wildfire characteristics ML model of FIG. 1 appropriately programmed to perform the process 200).

[0042] The system obtains a first set of data elements (205). In some implementations, the system can retrieve the data elements from external data sources. For example, the system can use SQL calls to retrieve data from one or more databases or calls to API provided by other data sources, for example, using web services. In some implementations, the system can retrieve data elements through an API. In some examples data elements in the set of data elements represent individual wildfire events.

[0043] The system determines wildfire associations (210). For example, and as described herein, the system identifies wildfire represented in the first set of data elements that are part of the same wildfire complex. Wildfire associations can be determined using heuristics and/or one or more wildfire association ML models. [0044] Heuristics can be expressed as a set of rules that define wildfire associations. In some examples, rules can be expressed as predicates and values, where the predicate defines a Boolean function, and if the Boolean is satisfied, the value is applied. For example, a rule can specify that, if a minimum distance between the perimeters of two wildfires is below a first threshold (e.g.. 1 kilometer or 2 kilometers), and the time between the end of one wildfire and the beginning of a second wildfire is below a second threshold (e.g., 1 week or 2 weeks), the wildfire events are determined to be associated as parts of a wildfire complex (i.e., the same wildfire complex). Note that this approach can be used to implement connected component analysis. For example, a rule can specify that, if the perimeters of two wildfires overlap and the wildfires overlap temporally, the wildfires can be considered part of the same wildfire complex.

[0045] In some implementations, a trained wildfire association ML model can be used to determine whether individual wildfire events are in the same wildfire complex. The wildfire association ML model can be any appropriate ML model (e.g., a gradient boosted decision tree, a random forest) that is configured to accept data relating to wildfires and to produce an indication of associations in terms of a common wildfire complex. The wildfire association ML model can be trained on data in which examples include features relating to the wildfire and a label indicates the wildfire complex to which the wildfire belongs. Example features can include, without limitation, the wildfire perimeter at one or more time points, the time the wildfire occurred, the wildfire duration, the wildfire speed, and so on.

[0046] When the sy stem determines that a wildfire belongs to a wildfire complex, the system aggregates the boundaries into a single wildfire complex and reevaluates the remaining wildfires to determine whether any additional wildfires should be aggregated into the wildfire complex. In some implementations, the boundary of a wildfire complex is union of the boundaries of the wildfires that are included in the complex. In some implementations, the boundary of a wildfire complex is defined by the outermost points of any wildfire in the wildfire complex. In some implementations, the system can smooth the boundary by, for example, creating a spline connecting the outermost points of the wildfires in the wildfire complex. The process of aggregating wildfires into wildfire complexes repeats until no further aggregation occurs (e.g.. no heuristics are satisfied).

[0047] FIG. 3 illustrates an example process of aggregating wildfires into a wildfire complex. In the example of FIG. 3, the system receives three region descriptions 310, 320, 330 of the same region. The region descriptions can describe the same point in time, or points in time that are within a configured threshold (e.g., 1 week or 2 weeks). A first region description 310 contains a first boundary 315 of a first wildfire, a second region description 320 contains a second boundary 325 of a second wildfire, and a third region description 330 contains a third boundary 335 of a third wildfire. In this example, the first boundary 315 and the second boundary 325 satisfy a heuristic for joining wildfires into a wildfire complex. That wildfire complex is shown by an aggregate boundary 345 in a region description 340.

[0048] In a next iteration, the aggregate boundary 345 and the third boundary 335 satisfy the heuristic for joining wildfires into a wildfire complex, which is show n by an aggregate boundary 355 for a wildfire complex that covers all three wildfires in a region description 350. In this example, the first boundary 315 and the third boundary 335 might not satisfy a heuristic for joining wildfires into a wildfire complex, illustrating the importance of the iterative approach.

[0049] Returning to FIG. 2, the system determines aggregate metrics (215). Rules can be defined for each metric defining how the metrics should be aggregated across wildfires in a w ildfire complex. Examples of rules can include computing a union, a mean, a maximum, minimum, and the like. For example, the size of a wildfire complex can be computed as the sum of the sizes of the wildfires that are included in the wildfire complex. The system can compute the speed of the wildfire complex, for example, as the mean speed of the wildfires included in the wildfire complex, or a mean w eighted by the size of each wildfire in the wildfire complex. In some examples, the start time of the w ildfire complex can be the minimum start time of any wildfire the wildfire complex, and the end time of the wildfire complex can be the maximum end time of any wildfire in the wildfire complex. In some examples, a duration of the wildfire complex can be computed as the maximum end time of any wildfire in the wildfire complex minus the minimum start time of any wildfire in the wildfire complex.

[0050] The system can train a ML model (220) using (i) the data elements obtained in operation 205 and (li) the aggregate metrics determined in operation 215. The training process for the ML model can depend on the ty pe of ML model used. For example, if the ML model is a neural network, the ML model can be trained using backpropagation where the labels on the data elements and aggregate metrics are used to compute a loss function and the loss function is used to selectively continue iterative training of the neural network.

[0051] Once the ML model is trained, the system can obtain features (250) to be used during inference time to generate predicted wildfire characteristics using the trained ML model. In some examples, features can be wildfire-related metrics organized as data layers representing an area of interest. The features can include any wildfire-related metrics as described above. Features can also include weather-related metrics such as wind direction, average temperature, and humidity. Features can include terrain-related features, weather-related features, etc. Features can be obtained from external data sources or through an API, as described above.

[0052] The system can determine wildfire associations (260). If the features are associated with multiple wildfires, the system can determine which wildfires are associated with the same wildfire complex. The system can determine wildfire associations by performing steps such as those described herein (e.g., those performed in operation 210). If the features are associated with a single wildfire, this operation can be omitted.

[0053] The system determines aggregate metrics (270). In cases where the system determines (260) that two or more wildfires were associated with a single wildfire complex, the system can compute aggregate metrics for the wildfire complex. The system perform steps analogous to those performed in operation 215. If the system determines that multiple wildfires were not associated with a single wildfire complex, this operation can be omitted.

[0054] The system can generate an input from at least a subset of the data elements and the aggregate metrics. The system can process the input using the trained ML model (255) that is configured to generate a wildfire risk prediction output that characterizes predicted future characteristics of a wildfire.

[0055] The result of processing the ML model is an output that includes predictions. For example, the output can include one or more data layers, and for each pixel in a data layer, a value indicating a property of the wildfire at the pixel. For example, if the selected ML model predicts wildfire speed, an output pixel reflects the predicted speed of the wildfire at the pixel in the event that the area represented by the pixel bums. [0056] The system can provide the predictions (260) included in the output. For example, the system can provide the predictions to a web server where the predictions are available over HTTP.

[0057] This specification uses the term ‘'configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed thereon software, firmware, hardware, or a combination thereof that, in operation, cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0058] Implementations of the subject matter and the functional operations described in this specification can be realized in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Implementations of the subject matter described in this specification can be implemented as one or more computer programs (i.e., one or more modules of computer program instructions) encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus. The computer storage medium can be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory^ device, or a combination of one or more of them. The program instructions can be encoded on an artificially-generated propagated signal (e.g., a machinegenerated electrical, optical, or electromagnetic signal) that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0059] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry (e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit)). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs (e.g.. code) that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them. [0060] A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document) in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network. [0061] In this specification the term “engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components, installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in some cases, multiple engines can be installed and running on the same computer or computers.

[0062] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry (e.g., a FPGA, an ASIC), or by a combination of special purpose logic circuitry and one or more programmed computers.

[0063] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memoiy or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data (e.g., magnetic, magneto-optical disks, or optical disks). However, a computer need not have such devices. Moreover, a computer can be embedded in another device (e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver), or a portable storage device (e.g., a universal serial bus (USB) flash drive) to name just a few.

[0064] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memoiy, media and memory devices, including by way of example semiconductor memory’ devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks.

[0065] [0001] To provide for interaction with a user, implementations of the subject matter described in this specification can be provisioned on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse, a trackball), by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device (e.g., a smartphone that is running a messaging application), and receiving responsive messages from the user in return. [0066] Data processing apparatus for implementing machine learning models can also include, for example, special-purpose hardware accelerator units for processing common and compute-intensive parts of machine learning training or production (i.e., inference, workloads).

[0067] Machine learning models can be implemented and deployed using a machine learning framework (e.g.. a TensorFlow framework, a Microsoft Cognitive Toolkit framework, an Apache Singa framework, an Apache MXNet framework). [0068] [0002]Implementations of the subject matter described in this specification can be realized in a computing system that includes a back-end component (e.g., as a data server) a middleware component (e g., an application server), and/or a front-end component (e.g., a client computer having a graphical user interface, a web browser, or an app through which a user can interact with implementations of the subject matter described in this specification, or any combination of one or more such back- end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN) and a wide area network (WAN) (e.g., the Internet).

[0069] [0003]The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some implementations, a server transmits data (e.g., an HTML page) to a user device (e.g., for purposes of displaying data to and receiving user input from a user interacting with the device), which acts as a client. Data generated at the user device (e.g., a result of the user interaction) can be received at the serv er from the device.

[0070] [0004]While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular implementations of particular inventions. Certain features that are described in this specification in the context of separate implementations can also be implemented in combination in a single implementation. Conversely, various features that are described in the context of a single implementation can also be implemented in multiple implementations separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a sub-combination or variation of a subcombination.

[0071] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the implementations described above should not be understood as requiring such separation in all implementations, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0072] Particular implementations of the subject matter have been described. Other implementations are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

Claims

CLAIMS What is claimed is:

1. A computer-implemented method for training a machine learning (ML) model to predict characteristics of adverse events, the method comprising: accepting, by a training system, a plurality of sets of data elements, wherein a first set of data elements describe a first property of a first adverse event, a second set of data elements describe the first property of a second adverse event, and a third set of data elements describe the first property' of a third adverse event; determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex, and in response: aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements into an aggregate set of data elements describing the first property' for the first adverse event complex; and training, by the training system, a ML model that is configured to generate predictions relating to properties of adverse events using at least a subset of the aggregate set of data elements and at least a subset of the third set of data elements.

2. The computer-implemented method of claim 1, wherein the first adverse event and the second adverse event are one or more of temporally related and spatially related.

3. The computer-implemented method of claim 1, wherein determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex comprises processing the first adverse event and the second adverse event through one or more heuristics.

4. The computer-implemented method of claim 1, wherein determining, by the training system, that the first adverse event and the second adverse event are associated wi th a first adverse event complex comprises processing the first adverse event and the second adverse event through an adverse event association model that generates a prediction indicating that the first adverse event and the second adverse event are associated with a first adverse event complex.

5. The computer-implemented method of claim 1, further comprising determining that the third set of data elements is not associated with the first aggregate adverse event.

6. The computer-implemented method of claim 1, wherein aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements comprises computing one or more of a union, a mean, a maximum, and a minimum.

7. The computer-implemented method of claim 1, further comprising: obtaining, by an execution system, a fourth set of data elements that describe a second property of a third adverse event and a fifth set of data elements that describe the second property of a fourth adverse event; determining, by the execution system, that the third adverse event and the fourth adverse event are associated with a second adverse event complex; determining, by the training system and using at least a subset of the fourth set of data elements and at least a subset of the fifth set of data elements, a sixth set of data describing a second aggregate property of the second adverse event complex; and processing an input that includes at least a subset of the sixth set of data and using the machine learning model to generate a prediction relating to the second property of the second wildfire complex.

8. A non-transitory computer storage medium encoded with a computer program, the computer program comprising instructions that when executed by a data processing apparatus cause the data processing apparatus to perform operations for training a machine learning (ML) model to predict characteristics of adverse events, the operations comprising: accepting, by a training system, a plurality of sets of data elements, wherein a first set of data elements describe a first property of a first adverse event, a second set of data elements describe the first property of a second adverse event, and a third set of data elements describe the first property of a third adverse event; determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex, and in response: aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements into an aggregate set of data elements describing the first property for the first adverse event complex; and training, by the training system, a ML model that is configured to generate predictions relating to properties of adverse events using at least a subset of the aggregate set of data elements and at least a subset of the third set of data elements.

9. The non-transitory computer storage medium of claim 8, wherein the first adverse event and the second adverse event are one or more of temporally related and spatially related.

10. The non-transitory computer storage medium of claim 8, wherein determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex comprises processing the first adverse event and the second adverse event through one or more heuristics.

11. The non-transitory computer storage medium of claim 8, wherein determining, by the training system, that the first adverse event and the second adverse event are associated wi th a first adverse event complex comprises processing the first adverse event and the second adverse event through an adverse event association model that generates a prediction indicating that the first adverse event and the second adverse event are associated with a first adverse event complex.

12. The non-transitory computer storage medium of claim 8, wherein operations further comprise determining that the third set of data elements is not associated with the first aggregate adverse event.

13. The non-transitory computer storage medium of claim 8, wherein aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements comprises computing one or more of a union, a mean, a maximum, and a minimum.

14. The non-transitory computer storage medium of claim 8, wherein operations further comprise: obtaining, by an execution system, a fourth set of data elements that describe a second property of a third adverse event and a fifth set of data elements that describe the second property of a fourth adverse event; determining, by the execution system, that the third adverse event and the fourth adverse event are associated with a second adverse event complex; determining, by the training system and using at least a subset of the fourth set of data elements and at least a subset of the fifth set of data elements, a sixth set of data describing a second aggregate property of the second adverse event complex; processing an input that includes at least a subset of the sixth set of data and using the machine learning model to generate a prediction relating to the second property of the second wildfire complex.

15. A system, comprising: one or more processors; and a computer-readable storage device coupled to the one or more processors and having instructions stored thereon which, when executed by the one or more processors, cause the one or more processors to perform operations for training a machine learning (ML) model to predict characteristics of adverse events, the operations comprising: accepting, by a training system, a plurality of sets of data elements, wherein a first set of data elements describe a first property of a first adverse event, a second set of data elements describe the first property of a second adverse event, and a third set of data elements describe the first property of a third adverse event; determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex, and in response: aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements into an aggregate set of data elements describing the first property' for the first adverse event complex; and training, by the training system, a ML model that is configured to generate predictions relating to properties of adverse events using at least a subset of the aggregate set of data elements and at least a subset of the third set of data elements.

16. The system of claim 15, wherein the first adverse event and the second adverse event are one or more of temporally⁷ related and spatially related.

17. The system of claim 15, wherein determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex comprises processing the first adverse event and the second adverse event through one or more heuristics.

18. The system of claim 15, wherein determining, by the training system, that the first adverse event and the second adverse event are associated with a first adverse event complex comprises processing the first adverse event and the second adverse event through an adverse event association model that generates a prediction indicating that the first adverse event and the second adverse event are associated with a first adverse event complex.

19. The system of claim 15, wherein operations further comprise determining that the third set of data elements is not associated with the first aggregate adverse event.

20. The system of claim 15, wherein aggregating at least a subset of the first set of data elements and at least a subset of the second set of data elements comprises computing one or more of a union, a mean, a maximum, and a minimum.