# US 12,001,938 B2 · Convolutional neural network for estimating a solar energy production indicator > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,001,938 B2 - Document ID: US12001938B2 - Status: Granted - Kind: B2 - Application number: 16/975,892 - Filed: 2019-02-14 - Published: 2021-01-07 - Granted: 2024-06-04 - Earliest claimed: 2018-03-07 - Assignees: Electricite de France - Applicants: Electricite De France - Inventors: Vincent Le Guen - Web page: https://patentgrep.com/patent/US12001938B2 ## CPC - G06N 3/044: Recurrent networks, e.g. Hopfield networks - G01W 1/10: Devices for predicting weather conditions (computers per se G06; display devices G09) - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/045: Combinations of networks - G06N 3/0455: Auto-encoder networks; Encoder-decoder networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/08: Learning methods - G06N 3/09: Supervised learning - G06N 3/096: Transfer learning - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G01W: METEOROLOGY (radar, sonar, lidar or analogous systems, designed for meteorological use G01S13/95, G01S15/88, G01S17/95) ## Abstract A method for estimating at least one energy production indicator of a solar energy production system, comprising: reception of at least one image obtained from a wide-angle image capture device, processing of said obtained image by at least one convolutional neural network comprising at least two layers respectively for the application of a convolution filter to the obtained image and for the estimation of an energy production indicator. ## Claims ### Claim 1 A method for estimating at least one energy production indicator of a solar energy production system of the photovoltaic cell type, comprising: reception of at least one image obtained from a wide-angle image capture device, the obtained image being a hemispherical image of the sky, processing of said obtained image by at least one convolutional neural network comprising at least two layers respectively for the application of a convolution filter to said obtained image and for the estimation of an energy production indicator; wherein a series of images is obtained from the wide-angle image capture device; wherein the images of the series are processed in a first processing by a recurrent convolutional neural network in order to obtain a context vector, said context vector being subjected to a second processing by a recurrent non-convolutional neural network in order to obtain the energy production indicator; and wherein the application of the convolution filter is implemented in the recurrent convolutional neural network and the estimation of the energy production indicator is implemented in the recurrent non-convolutional neural network. ### Claim 2 (depends on 1) The method according to claim 1, wherein the obtained image is captured by the wide-angle image capture device at a time t and wherein the energy production indicator is estimated at the time t corresponding to the time of capture of the obtained image. ### Claim 3 (depends on 1) The method according to claim 1, wherein the images of the series are processed sequentially by the recurrent convolutional neural network, and wherein the energy production indicator is predicted for a time t subsequent to the capture of the series of images. ### Claim 4 (depends on 3) The method according to claim 3, wherein the recurrent convolutional neural network is of the Long Short-Term Memory type, LSTM. ### Claim 5 (depends on 1) The method according to claim 1, wherein the recurrent non-convolutional neural network is of the Long Short-Term Memory type, LSTM. ### Claim 6 (depends on 1) The method according to claim 1, wherein the image obtained from the wide-angle image capture device is obtained by: application of an optical flow algorithm to the captured series of images in order to predict said obtained image, the obtained image corresponding to a time t subsequent to the captured series of images. ### Claim 7 (depends on 1) The method according to claim 1, wherein the image obtained from the wide-angle image capture device is obtained by: processing of the series of images by another convolutional neural network in order to predict said obtained image, the obtained image corresponding to a time t subsequent to the captured series of images. ### Claim 8 (depends on 1) The method according to claim 1, wherein the processing by the convolutional neural network comprises at least: a first layer for the application of convolution filters; a second layer for downsampling the output from the first layer. ### Claim 9 (depends on 8) The method according to claim 8, wherein the processing by the convolutional neural network comprises several pairs of successive first and second layers. ### Claim 10 (depends on 8) The method according to claim 8, wherein the layer for estimating the energy production indicator comprises processing by a neuron followed by application of a linear activation function. ### Claim 11 (depends on 1) The method according to claim 1, wherein the at least one energy production indicator comprises one of, or a combination of: a global horizontal irradiance parameter; a diffuse horizontal irradiance parameter; or a direct normal irradiance parameter. ### Claim 12 An estimation device for estimating at least one energy production indicator of a solar energy production system, comprising: a reception interface for receiving at least one image obtained from a wide-angle image capture device, wherein a series of images is obtained from the wide-angle image capture device, a processor configured for the processing of said captured image by at least one convolutional neural network comprising at least two layers respectively for the application of a convolution filter to said obtained image and for the estimation of an energy production indicator; wherein the processor is configured to process the images of the series in a first processing by a recurrent convolutional neural network in order to obtain a context vector, said context vector being subjected to a second processing by a recurrent non-convolutional neural network in order to obtain the energy production indicator; wherein the application of the convolution filter is implemented in the recurrent convolutional neural network and the estimation of the energy production indicator is implemented in the recurrent non-convolutional neural network. ### Claim 13 (depends on 12) A system comprising a wide-angle image capture device capable of communicating with the estimation device according to claim 12. ### Claim 14 (depends on 4) The method according to claim 4, wherein the recurrent non-convolutional neural network is of the Long Short-Term Memory type, LSTM.