Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/044—Recurrent networks, e.g. Hopfield networks
• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01W—METEOROLOGY
  • G01W1/00—Meteorology
  • G01W1/10—Devices for predicting weather conditions
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/044—Recurrent networks, e.g. Hopfield networks
  • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/045—Combinations of networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/045—Combinations of networks
  • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/0464—Convolutional networks [CNN, ConvNet]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/09—Supervised learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/096—Transfer learning

Definitions

• the present invention relates to the field of short-term estimation or forecasting of the production of electrical energy by photovoltaic cells.
• the invention relates to the estimation of current and / or future solar radiation exploitable for the production of energy and more particularly to the estimation of radiation parameters and / or the prediction of their evolution by analysis of images of cloud cover (cloudiness, which affects the exploitable radiation).
• An improved estimate of PV production represents an advantage for the generator and the grid operator to ensure a balance between consumption and production, which guarantees the safety and stability of the electrical system.
• these estimates or forecasts make it possible to optimize the operation of mixed systems combining photovoltaic production and storage systems (batteries) or other means of production (for example, the anticipation or the detection of a fall in production makes it possible to implement en route adapted batteries).
• the estimation of the solar radiation flux is generally carried out by a device comprising a pyranometer type sensor. Such a sensor is relatively expensive and is not adapted to the forecast.
• the estimation at a time t0 can be performed from a single acquired image.
• Estimation at a future time (forecast) of irradiation parameters may require the acquisition of a plurality of images.
• this plurality of images is pretreated to calculate characteristic parameters of the evolution of cloudiness.
• the data thus obtained are then processed to obtain the targeted parameters.
• the treatment may in particular be carried out by a neural network.
• the captured image (or images), which is distorted must be corrected beforehand (transformation to move to an orthonormal two-dimensional coordinate system).
• Such a correction requires a prior calibration of the image acquisition device using a test pattern, which calibration is preferably performed on the operating site of the sensor.
• a first aspect of the invention relates to a method for estimating at least one energy production indicator of a solar energy production system, comprising:
• processing of the image obtained by at least one convolutional neural network comprising at least two layers for respectively the application of a convolution filter to the image received and the estimation of an energy production indicator.
• the treatment by a convolutional network advantageously makes it possible to dispense with a prior calibration step of the image acquisition device or of the image transformation obtained by the device.
• the image obtained can be acquired by the wide-field image acquisition device at a time t and the energy production indicator can be estimated at the instant t corresponding to the instant of acquisition of the image obtained.
• the method according to this embodiment thus allows the real-time estimation of the energy production indicator.
• a series of images can be obtained from the wide-field image acquisition device, the images of the series can be processed sequentially by a recurrent convolutional neural network and the production indicator. of energy can be predicted for a time t subsequent to the acquisition of the series of images.
• the method according to this embodiment thus allows the estimation (or prediction) of future energy production indicators, which can advantageously make it possible to optimize the distribution of electrical energy in a network for example.
• the convolutional neural network may be a recurrent convolutional neural network
• the images of the series may be processed by a first treatment by the recurrent convolutional neural network in order to obtain a context vector
• the context vector being subjected to a second treatment by a network of non-neurons convolutional recursive to obtain the indicator of energy production.
• the application of the convolutional filter can be implemented in the recurrent convolutional neural network and the estimation of the energy production indicator can be implemented in the recurrent non-convolutional neuron network.
• Such an embodiment allows a precise estimate of the evolution of the indicator of solar energy production.
• the recurrent convolutional neural network may be Long Short Term Memory, LSTM.
• the recurrent convolutional neural network is also called an encoder network in an encoder / decoder type architecture.
• the recurrent non-convolutional neuron network may be of the Long Short Term Memory, LSTM type.
• the recurrent non-convolutional neural network is also called a decoder network in an encoder / decoder type architecture.
• the image obtained from the wide-field image acquisition device can be obtained by:
• the prediction of the future image can be performed upstream of the estimation device, which can thus be used both on directly acquired images or on predicted images.
• the image obtained from the wide-field image acquisition device can be obtained by:
• the prediction of the future image can be carried out upstream of the estimation device, which can thus be used both on directly acquired images or on predicted images.
• the treatment with the convolutional neural network comprises at least:
• the first layer makes it possible to duplicate the image as many times as there are characteristic plans, or "feature maps", and the second layer makes it possible to reduce the size of these plans.
• the treatment with the convolutional neural network can comprise several pairs of first and second successive layers.
• the layer for estimating the energy production indicator may include neuron processing followed by the application of a linear activation function.
• At least one energy production indicator may include one of, or a combination of:
• a second aspect of the invention relates to a computer program comprising a series of instructions which, when executed by a processor, implement the steps of a method according to the first aspect of the invention.
• a third aspect of the invention relates to a device for estimating at least one indicator of energy production of a solar energy production system, comprising:
• a processor configured for processing said image acquired by at least one convolutional neural network comprising at least two layers for respectively applying a convolution filter to said received image and estimating a production indicator; 'energy
• a fourth aspect of the invention relates to a system comprising a wide-field image acquisition device capable of communicating with an estimation device according to the third aspect of the invention.
• FIG. 1 shows a system according to one embodiment of the invention
• FIG. 2 shows a convolution layer applied to an image according to one embodiment of the invention
• FIG. 3 illustrates the different layers of a convolutional neural network according to one embodiment of the invention
• FIG. 4 illustrates the operation of a neuron of a neural network according to one embodiment of the invention
• FIG. 5 is a diagram illustrating the steps of an estimation method according to one embodiment of the invention.
• FIG. 6 schematically illustrates the principle of a recurrent network
• FIG. 7 illustrates a set of neural networks for the prediction of a solar energy production indicator according to one embodiment of the invention
• FIG. 8 illustrates the structure of an estimation device according to one embodiment of the invention.
• FIG. 1 illustrates a system for estimating at least one indicator of energy production of a solar energy production system according to one embodiment of the invention.
• the estimation system may preferentially be located near the solar energy production system.
• the solar power generation system is not shown in Figure 1.
• the system includes a wide field image acquisition device 100, such as a fisheye camera for example.
• Each image captured by the acquisition device 100 can be transmitted to an estimation device 102 according to the invention, or, according to other embodiments, to an optional prediction device 101, itself connected to the device of the invention. estimate 102.
• the estimation device 102 receives an image or images directly from the acquisition device 100 or from the prediction device, the received image or images are considered as obtained from the acquisition device 100.
• the estimation device 102 is configured to implement a method according to the invention, as described in more detail in the following.
• the invention is based on the use of a wide-field image acquisition device in combination with an estimation device integrating a convolutional type deep neural network, also called CNN for "convolutional neural network”. .
• the acquisition device 100 can be fixed or mobile and can acquire images at a given frequency, or on a one-off trigger. For example, the acquisition device 100 can acquire images every ten seconds. No restrictions are attached to the frame rate.
• a CNN type network may comprise a first processing layer, or convolution layer, which makes it possible to process the image obtained from the acquisition device 100 or the prediction device 101.
• the convolution layer can process the resulting image in groups of pixels.
• the CNN type network may comprise a plurality of (convolution) processing layers, and further includes a layer for extracting an estimate of at least one solar energy output indicator, such as an estimate of solar radiation.
• the indicator can be: a parameter representative of the electrical output from the power generation device such as a solar panel placed on the ground near the estimation system;
• the solar radiation by means of a radiation sensor, such as a SunTracker (of the Solys2 TM type, for example), or a pyranometer (of the SPN1 type for example).
• a radiation sensor such as a SunTracker (of the Solys2 TM type, for example), or a pyranometer (of the SPN1 type for example).
• the use of a CNN type network eliminates the need for a calibration of the acquisition device and the need to transform the digital image obtained from the acquisition device.
• FIG. 2 illustrates the principle of convolution and presents the result of the application of a convolutional filter 201 to an image 200 comprising pixels 203.
• the convolutional filter 201 is of size 3 * 3 and is therefore applied to each 3 * 3 pixel matrix of the image 200 to calculate a resulting pixel 204 of the resulting image 202.
• Multiple convolution filters can be applied to an image to duplicate the image into several feature maps or feature maps.
• FIG. 3 illustrates the different substeps corresponding to a step of applying a convolutional neuron network to an image.
• a convolutional network LeNet-5 generally used in the recognition of digit images (from 0 to 9, from which the ten neurons in the last layer), is illustrated in order to present the general principle of a convolutional neural network.
• Such digit images here have an input size of 32 * 32 pixels (black and white images).
• 5 * 5 pixel convolution filters are applied in order to obtain a series of characteristic planes, for example 6 feature plans, of size 28 * 28 pixels after the first convolution layer 301.
• the convolutional neural network can be constructed by learning from the processing of a set of images and by comparison with the data obtained by the radiation sensor or by the pyranometer, during an undescribed prior step. further below.
• a subsampling substep, or subsampling layer, referred to as “sub-sampling” or “max pooling”, may be applied to feature plans 301 to reduce the size of the image.
• the size of feature planes 301 is divided by 2 in order to obtain subsampled plans 302 of 14 * 14 pixels.
• Nonlinear activation functions can be introduced between the layers, for example after each convolution layer.
• An example of a non-linear function can be a function of the type "REctified Linear Unit" RELU.
• the sequence of several pairs of convolution and subsampling layers makes it possible to capture shapes with different levels of granularity, invariant by translation.
• a fully connected layer is a layer where each neuron is connected to all the neurons of the previous layer.
• a convolution layer does not belong to this type of layer
• Figure 3 shows a fully connected first layer to obtain 120 values, then a fully connected second layer to obtain 84 values and finally a fully connected third layer to obtain 10 values which are the ten outputs of the neural network.
• convolutional and which indicate, in the example considered, the recognition or not of a given figure).
• the fully connected layers make it possible to connect the captured forms to the information that is to be predicted with the convolutional neural network.
• a convolutional neural network In a convolutional neural network, a succession of several convolutional layers, followed by nonlinear activation functions, gradually builds an increasingly abstract representation of the input data. For example, if the convolutional neural network takes as input a car image, the first layer of convolution can detect lines or outlines. Then these elements are combined with each other by the following layers and a layer will detect the more abstract wheel concept, and the last layer will identify the concept of car.
• the principle of a convolutional neural network illustrated in FIG. 3 can advantageously be used in the context of estimating an energy production indicator.
• the last "fully connected" layer returns a single value and is an estimation layer of a power generation indicator.
• the use of a convolutional neural network makes it possible to overcome prerequisites such as camera calibration and pretreatments such as the definition of the indicators, the segmentation of the images, compared to the existing processes that do not use no convolutional neural network.
• Figure 4 illustrates a neuron of a processing layer of a neural network.
• a neural network is generally composed of a succession of layers each of which inputs the outputs of the previous layer.
• Each layer i is composed of N, neurons, taking their inputs on the N M neurons of the previous layer i-1.
• a neuron calculates a weighted sum of its inputs and then applies a nonlinear activation function, such as a sigmoid function or a hyperbolic tangent, to obtain an output value.
• the weights of the weights are parameters that are optimized during the learning phase.
• the illustrated neuron applies weights 402.1 - 402. n to the respective input values 401.1 - 401.n.
• the weighted values are then subjected to a combination function 403.
• a nonlinear activation function 405 is then applied based on a threshold value 404 to obtain the output value 406 of the neuron.
• the principle of a convolutional-type neural network is to apply transformations to an image by taking into account its two-dimensional structure by means of convolution filters.
• a convolutional filter of size 3 * 3 pixels applied to an image smooths the image by transforming each pixel by a linear combination of the values of its neighbors.
• the same filter is applied to the entire image, moving on the image from left to right and from top to bottom, for example, to calculate the output values.
• the choice of the architecture of the convolutional neural network can be made:
• High-performance neural network architectures can include millions of parameters to be trained, can include up to 150 layers (for the ResNet network for example) and can be trained for several weeks on graphic processing units, or " Graphics Processing Unit »GPU. Such pre-trained networks can thus be used to estimate an indicator of solar energy production.
• the present invention can provide for the use of a ResNet-type network whose parameters are driven on a learning basis (images obtained by a wide-field image device in association with measurements). an indicator of energy production, especially irradiation.
• This network can include up to 152 layers and more than 60 million parameters to be optimized.
• the invention has been implemented on a network of ResNet type. of 51 layers and comprising more than 23 million parameters.
• FIG. 5 is a diagram illustrating the steps of a method of estimating at least one indicator of solar energy production according to several embodiments of the invention.
• Steps 500 to 502 for learning the recurrent neural network are prior to steps 503 through 505.
• a series of images is acquired. Such a series of images can be acquired by a single image acquisition device or by several.
• measurements of the solar energy production indicator to be evaluated (or indicators) are carried out at a step 501. Each measurement corresponds temporally to the acquisition of one of the images of step 500.
• the parameters of the convolutional neural network are determined by learning from the data acquired in steps 500 and 501, and to determine one or more indicators of solar energy production.
• the convolutional neural network can be manually defined or can take advantage of an already existing architecture.
• the convolutional neural network whose parameters have been determined can be used to determine the solar energy output indicator.
• the estimation device 102 receives an image obtained from the wide-field image acquisition device.
• the image obtained is processed by the convolutional neural network resulting from steps 500 to 502, in order to obtain at step 505 the indicator of energy production.
• Steps 503 to 505 cover several embodiments of the invention, depending on whether the power generation indicator is an instantaneous (present) value or a predicted value for a future instant.
• the image obtained is acquired by the wide-field image acquisition device 100 at a time t and the image is processed by the convolutional neural network to estimate the production indicator.
• energy at time t corresponding to the moment of acquisition of the image obtained.
• the method estimates the indicator corresponding to a single image. The indicator is therefore representative of the solar radiation at the moment of acquisition of this image.
• a series of images is obtained from the wide-field image acquisition device 100, the images of the series being processed sequentially by a recurrent convolutional neural network.
• the energy production indicator can be predicted for a time t (strictly) subsequent to the acquisition of the series of images.
• the invention proposes the use of a convolutional neural network of recurrent type for processing a series of images.
• recurrent networks are a specialization for sequential data processing.
• a recurrent network does not take as input examples independent of each other, but series or sequences of examples, a sequence of examples being for example numbered from 1 to T (x1 , x2, ... xT).
• Figure 6 schematically illustrates the principle of a recurrent network.
• the network comprises an input layer with a vector x describing each example and cycles are introduced into the connections between the neurons s.
• the output of the network is noted o.
• the looping W making the recurrent network is detailed on the right part of FIG.
• Examples of a sequence x ti , x t , x t + i are presented one by one in the order of their numbering, the values s of the neurons then constituting a current state of the neural network.
• the output o t of the network for the neuron s t does not depend only on the input vector x t , but also on the state of the network at time t-1, and therefore of the one at all the previous instants.
• a variant of the so-called “simple” recurrent networks as described with reference to FIG. 6 is the LSTM network for "Long Short Term Memory”.
• An LSTM block or cell is composed of a state, an entry gate, an exit gate, and an oblivion gate.
• the state vector at time t represents the current state of the sequence of examples by knowing the past states, hence the name "memory”.
• the evolution of this memory is governed by the three doors, represented by weights to learn.
• LSTM networks were initially developed to address the "vanishing gradients" of "simple” recursive networks when applied to long sequences.
• the processing of a series of images according to the invention may involve the architecture represented with reference to FIG. 7.
• Such an architecture may be called encoder / decoder (or, in the literature, "sequence to sequence” used in particular in translation of text applications from one language to another).
• the architecture presents a first convolutional convolutional neural network 700 of the LSTM type which takes as input a series of images 704 acquired at different times, and processes each of them by a convolutional network (CNN in FIG. 7).
• Information from CNN convolutional neural networks is entered into a LSTM-type recurrent network.
• the state vector of the LSTM network, or context vector, referenced 701 is the representation of what the first recurrent convolutional network 700 has learned from the series of images (for example a vector describing how which clouds move).
• the first recurrent (convolutional) network 700 can thus be called a recurrent encoder network.
• the context vector 701 is used at the input of a second recurrent non-convolutional network 702 called the decoder recursive network.
• the recurrent decoder network 702 uses the context vector 701 to estimate the solar energy production indicators at future time steps to + H, t 0 + 2H and t 0 + 3H for example, t 0 being the instant of acquisition of the 704 series of images.
• Such a method thus consists in calculating the optical flow between successive images (for example, estimation of the displacement of the clouds) and propagation of this motion to predict a future image at t 0 + H, t 0 corresponding to the acquisition of the series of images. images.
• the estimation device 102 then applies the method presented above to the future image to obtain the estimate of the solar energy production indicator.
• a fourth embodiment aiming, as for the second mode and for the third mode, to estimate a future value of the indicator, the following steps are implemented:
• this fourth embodiment instead of calculating the optical flow to predict one or more future images, it is possible to use neural networks that directly predict the future image (s). This is the same neural network architecture as the one presented above, except that outputs are future images (not the indicator or indicators). The device 102 then processes the future images as described above.
• FIG. 8 illustrates the structure of the estimation device 102 according to one embodiment of the invention.
• the estimation device 102 comprises a random access memory 803 and a processor 802 for storing instructions enabling the implementation of the steps 503 to 505 of the method described above.
• the device also optionally includes a database 804 for storing data to be retained after the application of the method.
• the database 804 may further store the neural network (s), recurrent and non-recurrent, as well as their respective parameters.
• the estimation device 102 furthermore comprises an input interface 801 intended to receive the images coming from the prediction device 101 or from the wide-field image acquisition device 100, and an output interface 805 capable of providing data output the indicator of solar energy production.
• the method according to the invention has been implemented in a Resnet-type network with parameters pre-trained on the ImageNet database, the last layers of which have been modified to perform the estimation of the solar radiation.
• the network used takes as input a 224 * 224 * 3 color image (the images from the camera are centered and resized to that size). Indeed, three channels are used to represent the color image: red, green and blue.
• a first block called ResnetConv includes a Resnet network and the set of convolution layers. The last layers of the Resnet network used to perform a classification between 1000 classes of objects, they have been removed according to the invention in order to add layers instead of making the regression of the radiation, namely:
• a so-called "drop out” layer which is a regularization which consists in putting the input neurons at 0 with a probability of 0.5;
• the validation database consists of the images of the year 2016 on the same site.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Life Sciences & Earth Sciences (AREA)
• Computing Systems (AREA)
• Mathematical Physics (AREA)
• Biophysics (AREA)
• Computational Linguistics (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• General Health & Medical Sciences (AREA)
• Molecular Biology (AREA)
• Artificial Intelligence (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Biomedical Technology (AREA)
• Software Systems (AREA)
• Health & Medical Sciences (AREA)
• Environmental & Geological Engineering (AREA)
• Atmospheric Sciences (AREA)
• Biodiversity & Conservation Biology (AREA)
• Ecology (AREA)
• Environmental Sciences (AREA)
• Image Analysis (AREA)
• Photovoltaic Devices (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=62873435

Family Applications (1)

Country Status (4)

Families Citing this family (4)

Family Cites Families (2)

• 2018
  • 2018-03-07 FR FR1851971A patent/FR3078802B1/fr active Active
• 2019
  • 2019-02-14 WO PCT/EP2019/053670 patent/WO2019170384A1/fr not_active Ceased
  • 2019-02-14 US US16/975,892 patent/US12001938B2/en active Active
  • 2019-02-14 EP EP19703758.3A patent/EP3762872A1/de active Pending

Non-Patent Citations (3)

Also Published As

Similar Documents

Legal Events