WO2025262015A1

WO2025262015A1 - A method for short-term forecasting of a global horizontal irradiance value and associated training method

Info

Publication number: WO2025262015A1
Application number: PCT/EP2025/066843
Authority: WO
Inventors: Giulio Betti; Emanuele Giovanni Carlo OGLIARI; Maciej SAKWA; Sonia Leva
Original assignee: Nuovo Pignone Technologie SRL
Current assignee: Nuovo Pignone Technologie SRL
Priority date: 2024-06-17
Filing date: 2025-06-17
Publication date: 2025-12-26
Anticipated expiration: 2026-12-17

Abstract

The present disclosure relates to a computer-implemented method (200) of training a predictive model (10) for short-term forecasting of a Global Horizontal Irradiance, GHI, value at a reference location (21), the method (200) comprising: i) obtaining (201) a satellite image (S1-Sn) of cloud coverage over a ground area comprising said reference location (21); ii) performing (202) image processing of said satellite image (S1-Sn) to obtain a corresponding image of the cloud thickness over said ground area; iii) generating (203) a sun path image (P) by composing said image of the cloud thickness with a solar vector defining a path between said reference location (21) and a sun position at an acquisition time of said satellite image (S1-Sn), wherein said sun path image (P) is representative of a cloud occlusion of the sun along said solar vector; iv) providing (204) said image of the cloud thickness to a first feature extraction pipeline of said predictive model (10) to provide a first set of extracted features; v) providing (205) said sun path image (P) to a second feature extraction pipeline of said predictive model (10) to provide a second set of extracted features; vi) combining (206) said first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; and vii) inputting (207) said concatenated set of extracted features into a classification layer of the predictive model (10) for training on a relationship between said satellite image (S1-Sn) and said GHI value.

Description

A method for short-term forecasting of a Global Horizontal Irradiance value and associated training method

Description

TECHNICAL FIELD

[0001] The present disclosure concerns a method for training a predictive model for short-term forecasting of a Global Horizontal Irradiance (GHI) value at a reference location.

[0002] The subject matter disclosed herein also refers to a method for short-term forecasting of the GHI value at a reference location using the trained predictive model.

[0003] The subject matter disclosed herein also refers to a data processing apparatus comprising a processor to execute the above methods.

[0004] The subject matter disclosed herein also refers to a computer program product comprising one or more instructions that when executed by the processor, causes the processor to perform the above methods.

BACKGROUND ART

[0005] Electric power production from Renewable Energy Sources (RES) is characterized by intrinsic volatility due to reliance on weather conditions that often cause imbalances between power generation and demand.

[0006] Reliable method of predicting solar radiation are therefore crucial for maintaining the electric stability of the grid by estimating in advance the amount of photovoltaic (PV) based power generation avoiding curtailments due to limits of the grid capacity or due to electricity market constraints. PV power generation forecasting is a widely studied field and is based on the prediction of meteorological parameters, such as the Global Horizon Irradiation (GHI) due to its high correlation with the power produced by photovoltaic (PV) devices [1,2].

[0007] Currently, GHI forecasting is performed by using either analytical techniques, such as Numerical Weather Predictions (NWP), or by adopting a Machine Learning (ML) techniques, such as Deep Learning (DL) models [3], However, considering that GHI forecasting requires short-term forecasts, classical NWPs are often considered lacking in terms of both spatial and temporal resolution of the predictors. This leads to often reducing the NWP analytical equations to be used just with a supportive role during the forecast. Concerning the second approach, forecasting methods using ML techniques have been widely studied with confirmed results. Specifically, few different DL approaches have been tested by researchers based on the use of temporal weather data, satellite images or All-Sky Imager (ASI) images.

[0008] For example, N. Yogambal Jayalakshmi, et al “Novel multi-time scale deep learning algorithm for solar irradiance forecasting' Energies, 14(9), 2021, provides a Long Short-Term Memory (LSTM) to perform a time series forecast based on temporal weather data.

[0009] Yongju Son, et al “Cloud cover forecast based on correlation analysis on satellite images for short-term photovoltaic power forecasting' Sustainability, 14(8), 2022, discloses a differ approach in which satellite images are used for predictions of cloud coverage.

[0010] ML techniques are also applied to sky-dome images from an ASI camera and employed for irradiation forecasting. For example, Ariana Moncada et al “Deep learning to forecast solar irradiance using a six-month UTSA skyimager dataset", Energies, 11(8), 2018, describes a short-term forecasting technique in which sky-dome images from the ASI camera are used in combination with Artificial Neural Networks (ANN).

[0011] However, there is a need to provide a more reliable method of predicting solar radiation which allows to better balance electric grid capabilities and optimise electrical energy distribution.

SUMMARY

[0012] Certain aspects commensurate in scope with the originally claimed disclosure are summarized below. These aspects are not intended to limit the scope of the claimed disclosure, but rather these aspects are intended only to provide a brief summary of possible forms of the disclosure. Indeed, the full disclosure may encompass a variety of forms that may be similar to or different from the aspects set forth below. [0013] In one aspect, the subject matter disclosed herein is directed to a computer- implemented method of training a predictive model for short-term forecasting of the GHI value at a reference location, the method comprising: i) obtaining a satellite image of cloud coverage over a ground area comprising said reference location; ii) performing image processing of said satellite image to obtain a corresponding image of the cloud thickness over said ground area; iii) generating a sun path image by composing said image of the cloud thickness with a solar vector defining a path between said reference location and a sun position at an acquisition time of said satellite image, wherein said sun path image is representative of a cloud occlusion of the sun along said solar vector; iv) providing said image of the cloud thickness to a first feature extraction pipeline of said predictive model to provide a first set of extracted features; v) providing said sun path image to a second feature extraction pipeline of said predictive model to provide a second set of extracted features; vi) combining said first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; and vii) inputting said concatenated set of extracted features into a classification layer of the predictive model for training on a relationship between said satellite image and said GHI value.

[0014] In another aspect, the subject matter disclosed herein is directed to a computer-implemented method for short-term forecasting the GHI value at a reference location, the method comprising: obtaining a satellite image of cloud coverage over a ground area comprising a reference location; performing image processing of said satellite image to obtain a corresponding image of the cloud thickness over said ground area; generating a sun path image by composing said image of the cloud thickness with a solar vector defining a path between a sun position at an acquisition time of said satellite image and said reference location, wherein said sun path image is representative of a cloud occlusion of the sun along said solar vector; providing said image of the cloud thickness to a first feature extraction pipeline of a predictive model trained on a relationship between said satellite image and said GHI value, to provide a first set of extracted features; providing said sun path image to a second feature extraction pipeline of said predictive model to provide a second set of extracted features; combining said first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; inputting said concatenated set of extracted features into a classification layer of the predictive model to forecast said GHI value.

[0015] In a further aspect, the subject matter disclosed herein is directed to a data processing apparatus comprising a processor for carrying out the above provided methods.

[0016] In a further aspect, the subject matter disclosed herein is directed to an energy management system for an electric grid comprising one or more energy generators arranged to convert an energy source into electrical energy, said management system comprising the data processing apparatus, and a control unit being configured to manage said one or more energy generators and/or optimise electrical energy distribution to at least one of a plurality of electrical components based on said forecasted GHI value or the forecasted solar irradiation.

[0017] In a further aspect, the subject matter disclosed herein is directed to an electric grid comprising one or more energy generators and the energy management system.

[0018] In a further aspect, the subject matter disclosed herein is directed to a computer program product comprising instructions which, when the program is executed by a processor, cause the processor to carry out any one of the above methods.

[0019] In a further aspect, the subject matter disclosed herein is directed to a computer- readable medium having stored thereon the computer program product.

BRIEF DESCRIPTION OF THE DRAWINGS

[0020] A more complete appreciation of the disclosed embodiments of the invention and many of the attendant advantages thereof will be readily obtained as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings, wherein:

Fig. 1 is a schematic view of a data processing apparatus according to a preferred embodiment of the invention;

Fig. 2 is a flowchart of the method for short-term forecasting of a GHI value at a reference location, according to a preferred embodiment of the invention; Fig. 3 is a flowchart of the method for training a predictive model for shortterm forecasting of the GHI value at a reference location, according to a preferred embodiment of the invention;

Fig. 4 is a schematic view of the predictive model trained on a relationship between the satellite image and the GHI value, showing the concatenation layer that concatenates the output of respective feature extraction pipelines and details of the feature extraction pipelines;

Fig. 5 is a satellite image of cloud coverage over a ground area that comprises the reference location, according to a preferred embodiment of the invention;

Fig. 6 is a sun path image representative of a cloud occlusion of the sun along a solar vector, according to a preferred embodiment of the invention;

Fig. 7 is image of the sky modified to include the measured GHI value, according to a preferred embodiment of the invention;

Fig. 8 shows the result of a test carried out using satellite images of a clear sky;

Fig. 9 shows the result of a test carried out using satellite images of a partly cloudy sky;

Fig. 10 shows the result of a test carried out using satellite images obtained during a rainy day; and

Fig. 11 shows a first exemplary image of the cloud thickness over the ground area as obtained from the satellite image, in this image the one or more blocks of pixels (or image pixels) interfering with the solar vector are shown.

DETAILED DESCRIPTION OF EMBODIMENTS

[0021] The present disclosure concerns a method for short-term forecasting of a GHI value and an associated training method.

[0022] Reference is now made to Fig. 1 which shows a schematic view of a data processing apparatus 1 according to a preferred embodiment of the invention. The data processing apparatus 1 comprises a processor 12 for carrying out the method of training a predictive model 10 for short-term (i.e. intra-hour) forecasting of a GHI value at a reference location and/or the method of short-term forecasting the GHI value at the reference location, as will be described later with reference to respective Fig. 2 and Fig. 3.

[0023] The processor 12 may be a processor suitable for the execution of a computer program product including, by way of example, both general and special purpose microprocessor(s), and/or any one or more processors of any kind of digital computer. Generally, the processor 12 will receive instructions and data from a read only memory (ROM), or a random access memory (RAM), or an external memory or any combination thereof and execute the instruction to carry out the methods described below with reference to Fig. 2 and Fig. 3. The instructions may be part of a program product which can be stored on a computer-readable medium 14, such as a memory or storage internal or external to the processing apparatus 1, such as a remote server.

[0024] The processor 12 may be part of a data processing apparatus 1, which in turn may be part of an energy management system for an electric grid. The data processing apparatus 1 allows to optimize the energy use in an electric grid by coordinating the load appliances, battery storage, and solar panels on the basis of the forecasted GHI. The energy management system may be part of an electric grid which comprises one or more energy generators.

[0025] However, it is evident that the present invention should not be limited to the specific use in an electric grid herein described, in fact the energy management system can be used also in a combined heat and power grid CHP (also known as cogeneration system) or as part of a hybrid microgrid.

[0026] As such, the energy management system may comprise the data processing apparatus 1, one or more energy generators, such as solar panels, arranged to convert an energy source into electrical energy, and a control unit that is configured to manage the one or more energy generators and/or optimise electrical energy distribution to at least one of a plurality of electrical components based on the forecasted GHI value or a forecasted solar irradiation.

[0027] For example, during peak sun hours, the energy management system can prioritize the use of solar power and store the excess energy in the batteries. Then, when solar production decreases, e.g. when the forecasted GHI is below a threshold value, the energy management system can seamlessly switch to using the stored energy or grid power, minimizing reliance on the expensive grid and maximizing selfconsumption.

[0028] Reference is now made to Fig. 2 which shows a flowchart of the method 100 for short-term forecasting of a GHI value at a reference location, according to a preferred embodiment of the invention.

[0029] The method 100 comprises step 101 in which it is obtained a satellite image Sl-Sn of cloud coverage over a ground area comprising a reference location 21.

[0030] Fig. 1 shown an example of a satellite image Sl-Sn of cloud coverage over a ground area comprising a reference location. The satellite image Sl-Sn can be obtained from a storage, such as a memory of the data processing apparatus or a remote server, or directly from a geostationary satellite for weather forecasting. For example, satellite images may comprise atmosphere and cloud coverage images taken by geostationary satellites designed for weather forecasting and early detection of severe weather conditions.

[0031] Fig. 5 shows an example of the satellite image Sl-Sn which shows the cloud coverage in the surrounding of the reference location 21. Each image pixel of the satellite image Sl-Sn may comprise a cloud top height value. The cloud top height value may be associated with a cloud category, for example as indicated in the metadata of the satellite image Sl-Sn. The category type can be classified according to different classes given by the satellite image provided. For example, satellite images can be divided cloud typologies into 19 different classes. In this example, they are regrouped into 3 specified groups (low, medium, high).

[0032] At step 102 of the method 100, the satellite image Sl-Sn is processed to obtain a corresponding image of the cloud thickness over the ground area. As such, the image of the cloud thickness over the ground area may be a three-dimensional representation of the cloud coverage, as obtained the satellite image Sl-Sn, over the ground area. The processing step may comprise generating the image of the cloud thickness based on the cloud top height value and the associated cloud category. For example, for each image pixel of the satellite image Sl-Sn, step 102 may include the following sub-steps: i) defining a cloud bottom height value based on the associated could category; ii) calculating a cloud thickness value by subtracting the cloud bottom height value from the cloud top height value; and iii) associating the cloud thickness value to a corresponding pixel of the image of the cloud thickness.

[0033] For example, for each cloud, the expected bottom height can be defined based on the respective category type defined according to the typical heights found in the literature. Therefore, the thickness can be a result of subtraction between the top and bottom heights, where the top is drawn out from the original satellite image and the bottom depends on the cloud category.

[0034] However, it is evident that the present invention should not be limited to the specific method of calculating a cloud thickness value herein described.

[0035] At step 103 of the method it is generated a sun path image P by composing the image of the cloud thickness with a solar vector defining a path between a sun position at an acquisition time of the satellite image Sl-Sn and the reference location 21. In other words, the solar vector represents the path of the light from the sun position to the reference location 21 and the sun path image P represents the dynamic solar occlusion, derived from the single satellite image.

[0036] Fig. 11 shows an exemplary image of the cloud thickness over the ground area as obtained in step 102. The clouds are represented as blocks of pixels within the image. In this figure the solar vector connects the sun position with the reference location 21, at any given time, providing a time dependent information. As shown in Fig. 11, the combination of the solar vector and image of the cloud thickness over the ground area (obtained in step 102) results in one or more blocks of pixels (or image pixels) interfering with the solar vector. Specifically, in this example, the blocks of pixels interfering with the solar vector are associated with type B clouds, whereas the remaining blocks of pixels, that is those associated with type A clouds, do not interfere with the solar vector. As such, it is possible to determine which of the one or more blocks of pixels (or pixels) in image of the cloud thickness results in solar radiation attenuation, according to the respective cloud category.

[0037] Fig. 6 shows a sun path image P representative of a cloud occlusion of the sun along a solar vector, according to a preferred embodiment of the invention. The role of the solar vector is to connect the sun position found at a given time (i.e. the image acquisition time) with the reference location 21, providing a time-dependent information in the obtained sun path image P to be fed to the predictive model 10.

[0038] The solar vector can be used in a series of orientation tests to identify the clouds that lay directly on the path between the sun and the reference location 21. For example, the solar vector can be calculated as a projection on the satellite image Sl- Sn of a line between the sun position at the acquisition time and the reference location 21. As such, the sun path image P comprises one or more image pixels of the image of the cloud thickness that intersect with the solar vector.

[0039] With reference to Fig. 2 the method 100 further comprises step 104 in which the image of the cloud thickness is fed to a first feature extraction pipeline of a predictive model 10 trained on a relationship between the satellite image Sl-Sn and the GHI value. The predictive model 10 may be trained as described later with reference to Fig. 3 or may be any model that is trained on a relationship between the satellite image Sl-Sn and the GHI value.

[0040] At step 105, the sun path image P is fed to a second feature extraction pipeline of the predictive model 10 to provide a second set of extracted features. The first set of extracted features is combined with the second set of extracted features at step 106 to generate a concatenated set of extracted features.

[0041] Finally, at step 107 the concatenated set of extracted features is inputted into a classification layer of the predictive model 10 to forecast the GHI value.

[0042] In this way it is possible to enable managing of one or more energy generators and/or optimization of electrical energy distribution to at least one of a plurality of electrical components, such as load appliances and/or battery storage, based on the forecasted GHI value (or forecasted solar irradiation). For example, the control unit of the energy management system may be configured to estimate the photovoltaic (PV) power output value and cause, by outputting instructions, at least one device of an electric grid, or power distribution network, to modify operation based on the estimated PV power output value. In some embodiments, the control unit may also cause at least one device of the electric grid to modify operation based on the forecasted GHI value (or forecasted solar irradiation), for example by outputting instructions to a substation.

[0043] In one example, the control unit may predict based on the forecasted GHI the corresponding photovoltaic production of one or more energy generators and control the charging and discharging of Battery Energy Storage Systems (BESS) according to the corresponding predicted photovoltaic production allowing for BESS optimization. Specifically, the control unit may instruct BESS to store excess energy produced during peak irradiance periods and then release it when the photovoltaic production decreases or the grid load increases. This helps reduce stress on the grid and maximize self-consumption.

[0044] In another example, the method may support grid frequency regulation reducing predicted sudden variations in photovoltaic production due to cloud passages that can cause fluctuations in the grid frequency. Predicting these fluctuations allows the control unit to proactively activate balancing resources (such as power reserves, modulation of other plants, or BESS) to maintain the frequency of the electrical grid within safety limits and ensure electric grid stability.

[0045] It is to be understood, however, that the use of the forecasted GHI value or solar irradiance value should not be limited to the specific applications described herein, which are provided solely by way of example.

[0046] Fig. 4 shows a schematic view of the predictive model 10 trained on a relationship between the satellite image and the GHI value, showing details of the feature extraction pipelines and a concatenation layer that is configured to concatenate the output generated by the respective feature extraction pipelines. The predictive model 10 shown in the figure is a first Convolutional Neural Network Model (CNN).

[0047] Each of the feature extraction pipelines can be broadly described as a three- block procedure where each block is composed of two convolution layers followed by a 2D max pooling layer. Both pipelines may be identical to each other with the same number and order of layers, number and size of neurons, as well as activation functions.

[0048] As shown in the Fig. 4, each convolution layer is based on 3x3 kernels with stride equal to 1, and tanh activation function. However, the number of filters doubles with each block starting from 64 up to 256. In the first block, the pooling operation is omitted due to the small size of the image; however, its removal can improve the feature extraction process and the predictive model 10 performance. The length of the twin feature extraction pipelines may be limited by the small size of the initial image (15x15 pixels).

[0049] After feature extraction, the two pipelines are concatenated in the concatenation layer and a regression prediction can be performed. For that goal, 3 additional hidden layers can be used with a decreasing number of units (256, 128, and 64 respectively) and a Rectified Linear Unit (ReLU) activation function. For model regularization, batch normalization and 20% dropout may be added between each hidden layer.

[0050] The predictive model 10 can be trained to convergence, with early stopping enabled to avoid overfitting. The training of the model will be described hereinbelow with reference to Fig. 3. The procedure may be performed using, for example, an Adam optimizer with the initial learning rate equal to 0.0003 which is reduced by a factor of 0.1 when a fitting plateau is reached. Moreover, in a further example to achieve better performance and further reduce the chance of overfitting, simple data augmentation can be performed during training with random image rotation (e.g. up to 5%), random horizontal flip, and random zoom (e.g. up to 110%).

[0051] Fig. 3 show a flowchart of the method 200 for training a predictive model 10 for short-term forecasting of the GHI value at a reference location 21, according to a preferred embodiment of the invention.

[0052] The method 200 allows to train the predictive model 10 using a dataset of satellite images Sl-Sn and comprises the following steps which can be executed for each of a plurality of satellite images Sl-Sn of cloud coverage over the ground area comprising the reference location 21. In particular, the method 200 comprises the following steps: i. obtaining 201 a satellite image Sl-Sn of cloud coverage over a ground area comprising the reference location 21, said satellite image Sl-Sn may be associated with a GHI value; ii. performing 202 image processing of the satellite image Sl-Sn to obtain a corresponding image of the cloud thickness over the ground area; iii. generating 203 a sun path image P by composing the image of the cloud thickness with a solar vector defining a path between the reference location 21 and a sun position at an acquisition time of the satellite image Sl-Sn, wherein the sun path image P is representative of a cloud occlusion of the sun along the solar vector; iv. providing 204 the image of the cloud thickness to a first feature extraction pipeline of the predictive model 10 to provide a first set of extracted features; v. providing 205 the sun path image P to a second feature extraction pipeline of the predictive model 10 to provide a second set of extracted features; vi. combining 206 the first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; and vii. inputting 207 the concatenated set of extracted features into a classification layer of the predictive model 10 for training on a relationship between the satellite image Sl-Sn and the GHI value.

[0053] Reference is now made to Fig. 7 which shows image of the sky K, such as a sky-dome image from an ASI camera modified, modified to include the measured GHI value for improved irradiation forecasting, according to a preferred embodiment of the invention.

[0054] The image of the sky K can be used to provide a short-term forecast of the solar irradiation using a system, such as a Convolutional Neural Network Model (CNN), that is trained on a relationship between the modified image of the sky K and the GHI value.

[0055] In particular, with continuous reference to Fig. 2, the method 100 may further comprises the following steps: a. obtaining image data from an image of the sky K captured by an image capture device; b. obtaining a current GHI value being indicative of a solar irradiation at an acquisition time of the image of the sky K; c. embedding the current GHI value into one or more pixels of the image of the sky to obtain a modified image of the sky K; and d. providing the modified image of the sky K to a system trained on a relationship between the modified image of the sky K and the GHI value to produce a short-term forecast of the solar irradiation.

[0056] In step b. the current GHI value is contextual to the image of the sky K, accordingly, the current GHI value is a measure of the solar irradiation at the reference location 21. For example, the current GHI value may be obtained using a local weather station, such as weather station located at the reference location 21 or nearby the reference location 21.

[0057] The steps a-d can be performed in response to a determination that the minimum acquisition time is above a predetermined time threshold, such as a threshold value within the interval of 10-15 minutes. The minimum acquisition time for a received set of satellite images Sl-Sn can be determined based on the acquisition time of each of the satellite image Sl-Sn, for example using all sky image metadata that provide information of the acquisition time.

[0058] The irradiation forecasting method based on steps a-d is particularly suited for very short-term solar radiation forecasting, e.g. under 15 minutes time-horizon, because it can overcome the inherent delay for a geostationary satellite to acquire and transmit the image to the data processing apparatus 1, allowing more accurate realtime predictions for very short-term solar radiation forecasting, for example under the 15-minute threshold.

[0059] Moreover, the method 100 may be configured to switch between very shortterm solar radiation forecasting using the image of the sky K and short-term solar radiation forecasting using the satellite image described with reference to Fig. 2, based on the comparison of the minimum acquisition time with predetermined time threshold.

[0060] For the purposes of this document, the term "short term" refers to any time period that is within a single hour (i.e. intra-hour interval), for example the term "short term" may comprise an intra-hour interval of any duration greater than 15 minutes.

Test Case

[0061] The following test case is based on a CNN-based method for short-term forecasting of the GHI value with a 15-minute time horizon to facilitate control and optimization of PV panel power generation.

[0062] The tested method is based on atmosphere and cloud coverage images taken by geostationary satellites designed for weather forecasting and early detection of severe weather conditions. The performance of the forecasting procedure is evaluated.

[0063] For this purpose, a robust dataset is used that is composed of two parts. Both are similar in structure and comprise of a series of satellite images of the cloud coverage in the surroundings of the reference location (in this example, latitude: 45.50°N; longitude: 9.16°E). The images are subdivided into pixels that represent an area of roughly 4 km by 4 km each. Moreover, for each pixel, two types of data are specified: the recorded cloud top height (in meters) paired with the cloud category (one of 19 cloud types).

[0064] The images in the selected periods are provided in 15-minute temporal resolution. The first portion of the dataset is acquired in the period between the beginning of September 2019 and the end of April 2020. These images have a size of 7x7 pixels corresponding to roughly 32 km by 32 km area and are centred on the reference location.

[0065] The second part has been acquired starting from the beginning of April 2023 to the middle of November 2023 and the images are twice the size at 15x15 pixels. With that, they cover an area of roughly 69 km by 69 km in the surroundings of the reference location. As the images differ in size, the first step of pre-processing was performed by upscaling the older images to 15x15 pixels. The summary of the information on the two parts of the dataset can be seen in table 1 below.

Table 1

[0066] The images represent the main dataset used for training, validation, and testing of the forecasting model. Combining the two parts, in total there are 37344 images. However, this amount is reduced to 10375 as sun irradiation cannot be forecasted at night (with a constant 0 value) and only images where the sun elevation is above 20° are used by the model. Each image is associated with corresponding real (weather sensor measured at ground level) and theoretical (clear-sky model [4]) GHI values. Their ratio is defined as Clear Sky Index (CSI) and is used as the model target label.

[0067] The data has been divided into train, validation, and test datasets with 76-19-5 proportions. For evaluation, 5 days from the test dataset have been selected that represent different weather conditions. This has been done to estimate the performance of the model with respect to the amount of cloud cover found in the images and the expected power generation of the PV modules. In detail, the days are:

- Test day 1 - 19/08/2023 - mostly clear sky

- Test day 2 - 23/08/2023 - mostly clear sky

- Test day 3 - 26/08/2023 - partly cloudy

- Test day 4 - 27/08/2023 - rain (very cloudy)

- Test day 5 - 28/08/2023 - cloudy

[0068] To assess the performance of the designed CNN model it has been compared against a benchmark. For this assessment, persistence model was selected as being the most typical benchmark for very short-term forecasts. The persistence is a highly accurate and simple benchmark, a basic prediction scheme, that simply assumes preservation of the temporal variable and predicts the last observed quantity as the model output. The performance was compared Forecast Skill (FS) defined as: where RMSE_pred and RMSEpers are the Root Mean Squared Error of the model and persistence respectively, and the RMSE is defined as: [0069] The model performance in terms of forecast skill is reported in Table 2, as follows.

Table 2

[0070] In the evaluation of the designed model, distinct trends emerge across varying meteorological conditions. Starting from sunny days (as shown in Fig. 8), the model exhibits precision in closely mirroring the recorded irradiance. This is made evident through the achieved FS with values highly above 0. The addition of the temporal data in the form of the solar vector (solar azimuth) through a secondary input was particularly helpful in this regard. However, in the case of a 15-minute forecast, the persistence error is relatively high and would decrease significantly for lower time horizons. Therefore, for very short-term nowcasting (5-, 10-minute) the model performance would have to be enhanced, for example by switching to an all sky image forecasting method during rainy conditions. [0071] Regarding overcast and partly cloudy conditions, the model adapted to the overarching diurnal pattern of the GHI throughout the day. Nonetheless, issues arose with unforeseen spikes attributed most probably to passing clouds, notably around the 40th and 80th minute of the recording (as can be seen on Fig. 9). Despite these discrepancies, the model maintained a sufficient performance indicated by FS slightly above 0.

[0072] In more challenging weather scenarios, such as rainfall (seen on Fig. 10) the model performance is impacted due to inherent volatility in the measurement, with the FS value dropping to below 0 regions. Nonetheless, the model can still generally follow the daily pattern of ascent and descent, indicating space for potential performance improvement with further study. However, it is important to note that such days might have lesser significance to the general power prediction, given the diminished expected power output of PV plants during rainy conditions.

[0073] A first advantage of the present disclosure is to provide a more. The proposed approach achieves very high performance, in particular during in clear-sky weather conditions, with the diurnal pattern closely followed by the developed forecasting model.

[0074] A second advantage is that the present solution allows better balance electric grid capabilities, which result in an optimised electrical energy distribution to at least one of a plurality of electrical components based on the forecasted GHI value and/or the forecasted solar irradiation. For example, the method allows to estimate in advance the amount of irradiance avoiding curtailments due to limits of the grid capacity or due to electricity market constraints. In summary, the short-term forecasting of solar irradiance, based on the irradiance value, is a crucial tool for the intelligent and reliable management of the modem electricity grid, characterized by an increasing integration of intermittent renewable sources. A third advantage of the present disclosure to provide an improved method for training a forecast model by using additional information content of the modified set of images, such as by using sun path images together with images of the cloud thickness or by embedding the measured GHI value into all sky camera images.

[0075] A fourth advantage is that the present solution allows to seamlessly switch between very short-term solar radiation forecasting using the image of the sky K and short-term solar radiation forecasting using the satellite image, overcoming the inherent delay for a geostationary satellite to acquire and transmit the image to the data processing apparatus 1 when performing very short-term solar forecasting and/or improving the forecasting during rainy conditions.

[0076] While aspects of the invention have been described in terms of various specific embodiments, it will be apparent to those of ordinary skill in the art that many modifications, changes, and omissions are possible without departing form the spirit and scope of the claims. In addition, unless specified otherwise herein, the order or sequence of any process or method steps may be varied or re-sequenced according to alternative embodiments.

[0077] The subject matter described herein can be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structural means disclosed in this specification and structural equivalents thereof, or in combinations of them. The subject matter described herein can be implemented as one or more computer program products, such as one or more computer programs tangibly embodied in an information carrier (e.g., in a machine-readable storage device), or embodied in a propagated signal, for execution by, or to control the operation of, data processing apparatus (e.g., a programmable processor, a computer, or multiple computers). A computer program (also known as a program, software, software application, or code) can be written in any form of programming language, including compiled or interpreted 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 computer program does not necessarily correspond to a file. A program can be stored in a portion of a file that holds other programs or data, 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 at one site or distributed across multiple sites and interconnected by a communication network.

[0078] The processes and logic flows described in this specification, including the method steps of the subject matter described herein, can be performed by one or more programmable processors executing one or more computer programs to perform functions of the subject matter described herein by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus of the subject matter described herein can be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).

[0079] Processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processor of any kind of digital computer. Generally, a processor will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. 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. Information carriers suitable for embodying computer program instructions and data include all forms of non-volatile memory, 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 optical disks (e.g., CD and DVD disks). The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.

[0080] To provide for interaction with a user, the subject matter described herein can be implemented 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 or 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, or tactile feedback), and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0081] The techniques described herein can be implemented using one or more modules. As used herein, the term “module” refers to computing software, firmware, hardware, and/or various combinations thereof. At a minimum, however, modules are not to be interpreted as software that is not implemented on hardware, firmware, or recorded on a non-transitory processor readable recordable storage medium (i.e., modules are not software per se). Indeed “module” is to be interpreted to always include at least some physical, non-transitory hardware such as a part of a processor or computer. Two different modules can share the same physical hardware (e.g., two different modules can use the same processor and network interface). The modules described herein can be combined, integrated, separated, and/or duplicated to support various applications. Also, a function described herein as being performed at a particular module can be performed at one or more other modules and/or by one or more other devices instead of or in addition to the function performed at the particular module. Further, the modules can be implemented across multiple devices and/or other components local or remote to one another. Additionally, the modules can be moved from one device and added to another device, and/or can be included in both devices.

[0082] The subj ect matter described herein can be implemented in a computing system that includes a back-end component (e.g., a data server), a middleware component (e.g., an application server), or a front-end component (e.g., a client computer having a graphical user interface or a web browser through which a user can interact with an implementation of the subject matter described herein), or any combination of such back-end, middleware, and 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.

[0083] Reference has been made in detail to the embodiments of the disclosure, one or more examples of which are illustrated in the drawings. Each example is provided by way of explanation of the disclosure, not limitation of the disclosure. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made in the present disclosure without departing from the scope or spirit of the disclosure. Reference throughout the specification to "one embodiment" or "an embodiment" or “some embodiments” means that the particular feature, structure or characteristic described in connection with an embodiment is included in at least one embodiment of the subject matter disclosed. Thus, the appearance of the phrase "in one embodiment" or "in an embodiment" or "in some embodiments" in various places throughout the specification is not necessarily referring to the same embodiment s). Further, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

[0084] When elements of various embodiments are introduced, the articles “a”, “an”, “the”, and “said” are intended to mean that there are one or more of the elements. The terms “comprising”, “including”, and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements.

References

[1] Razin Ahmed, Victor Sreeram, Y Mishra, and MD Arif. A review and evaluation of the state-of-the-art in pv solar power forecasting: Techniques and optimization. Renewable and Sustainable Energy Reviews, 124: 109792, 2020.

[2] Rial A Rajagukguk, Raden AA Ramadhan, and Hyun- Jin Lee. A review on deep learning models for forecasting time series data of solar irradiance and photovoltaic power. Energies, 13(24):6623, 2020.

[3] J. Antonanzas, N. Osorio, R. Escobar, R. Urraca, F.J. Martinez de Pison, and F. Antonanzas-Torres. Review of photovoltaic power forecasting. Solar Energy, 136:78- 111, 2016.

[4] Pierre Ineichen. A broadband simplified version of the Solis clear sky model. University of Geneva. Centre Universitaire D’Etude des problem de FEnergie, 2007.

Claims

1. A computer-implemented method (200) of training a predictive model (10) for short-term forecasting of a Global Horizontal Irradiance, GHI, value at a reference location (21) to enable managing of one or more energy generators and/or optimization of electrical energy distribution to at least one of a plurality of electrical components based on said forecasted GHI value, the method (200) comprising: i) obtaining (201) a satellite image (Sl-Sn) of cloud coverage over a ground area comprising said reference location (21); ii) performing (202) image processing of said satellite image (Sl-Sn) to obtain a corresponding image of the cloud thickness over said ground area; iii) generating (203) a sun path image (P) by composing said image of the cloud thickness with a solar vector defining a path between said reference location (21) and a sun position at an acquisition time of said satellite image (Sl-Sn), wherein said sun path image (P) is representative of a cloud occlusion of the sun along said solar vector; iv) providing (204) said image of the cloud thickness to a first feature extraction pipeline of said predictive model (10) to provide a first set of extracted features; v) providing (205) said sun path image (P) to a second feature extraction pipeline of said predictive model (10) to provide a second set of extracted features; vi) combining (206) said first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; and vii) inputting (207) said concatenated set of extracted features into a classification layer of the predictive model (10) for training on a relationship between said satellite image (Sl-Sn) and said GHI value.

2. The method (200) according to claim 1, further comprising performing steps (i)-(vii) for each of a plurality of satellite images (Sl-Sn) of cloud coverage over the ground area comprising said reference location (21).

3. The method (200) according to claim 1 or 2, wherein obtaining (201) a satellite image (Sl-Sn) of cloud coverage comprises obtaining said satellite image (Sl-Sn) from a storage or from a geostationary satellite for weather forecasting.

4. The method (200) according to any one of the preceding claims, wherein each image pixel of the satellite image (Sl-Sn) comprises a cloud top height value, said cloud top height value being associated with a cloud category, and wherein performing (202) image processing to obtain a corresponding image of the cloud thickness over said ground area comprises generating the image of the cloud thickness based on said cloud top height value and said associated cloud category.

5. The method (200) according to the preceding claim, wherein performing (202) image processing to obtain a corresponding image of the cloud thickness over said ground area comprises, for each image pixel of the satellite image (Sl-Sn): defining a cloud bottom height value based on said associated could category; calculating a cloud thickness value by subtracting said cloud bottom height value from said cloud top height value; and associating said a cloud thickness value to a corresponding pixel of said image of the cloud thickness.

6. The method (200) according to any one of the preceding claims, further comprising calculating said solar vector as a projection on said satellite image (Sl-Sn) of a line between the sun position at said acquisition time and the reference location (21).

7. The method (200) according to any one of the preceding claims, wherein said sun path image (P) comprising one or more image pixels of the image of the cloud thickness that intersect with said solar vector.

8. The method (200) according to any one of the preceding claims, wherein said predictive model (10) is a first Convolutional Neural Network Model, CNN.

9. A computer-implemented method (100) for short-term forecasting a Global Horizontal Irradiance, GHI, value at a reference location (21), the method (100) comprising: obtaining (101) a satellite image (Sl-Sn) of cloud coverage over a ground area comprising a reference location (21); performing (102) image processing of said satellite image (Sl-Sn) to obtain a corresponding image of the cloud thickness over said ground area; generating (103) a sun path image (P) by composing said image of the cloud thickness with a solar vector defining a path between a sun position at an acquisition time of said satellite image (Sl-Sn) and said reference location (21), wherein said sun path image (P) is representative of a cloud occlusion of the sun along said solar vector; providing (104) said image of the cloud thickness to a first feature extraction pipeline of a predictive model (10) trained on a relationship between said satellite image (Sl-Sn) and said GHI value, to provide a first set of extracted features; providing (105) said sun path image (P) to a second feature extraction pipeline of said predictive model (10) to provide a second set of extracted features; combining (106) said first set of extracted features with the second set of extracted features to generate a concatenated set of extracted features; inputting (107) said concatenated set of extracted features into a classification layer of the predictive model (10) to forecast said GHI value to enable managing of one or more energy generators and/or optimization of electrical energy distribution to at least one of a plurality of electrical components based on said forecasted GHI value.

10. The method (100) of claim 9, further comprising a) obtaining image data from an image of the sky (K) captured by an image capture device; b) obtaining a current GHI value being indicative of a solar irradiation at an acquisition time of said image of the sky (K); c) embedding said current GHI value into one or more pixels of said image of the sky to obtain a modified image of the sky (K); and d) providing said modified image of the sky (K) to a system trained on a relationship between the modified image of the sky (K) and the GHI value to produce a short-term forecast of the solar irradiation to enable managing of one or more energy generators and/or optimization of electrical energy distribution to at least one of a plurality of electrical components based on said forecasted solar irradiation.

11. The method (100) according to the preceding claim, further comprising determining the minimum acquisition time for a received set of satellite images (Sl-Sn) based on the acquisition time of each of said satellite image (Sl-Sn); and in response to the minimum acquisition time being above a predetermined time threshold, performing the steps (a)-(d).

12. The method (100) according to the preceding claim, wherein said predetermined threshold is a value within the interval of 10-15 minutes.

13. The method (100) according to the preceding claim, further comprising in response to determining the satellite images (Sl-Sn) are indicative of rainy condition, performing the steps (a)-(d).

14. The method (100) according to any one of claims 10-13, wherein said system comprises a second Convolutional Neural Network Model, CNN.

15. A data processing apparatus (1) comprising a processor (12) for carrying out the method (200) of any one of claims 1-8 or the method (100) of any one of claims 9- 14.

16. An energy management system for an electric grid comprising one or more energy generators arranged to convert an energy source into electrical energy, said management system comprising: the data processing apparatus (1) of claim 15, and a control unit being configured to manage said one or more energy generators and/or optimise electrical energy distribution to at least one of a plurality of electrical components based on said forecasted GHI value or said forecasted solar irradiation.

17. An electric grid comprising one or more energy generators and the energy management system of claim 16.

18. A computer program product comprising instructions which, when the program is executed by a processor (12), cause the processor (12) to carry out the method of any one of claims 1-8 or the method (100) of any one of claims 9-14.

19. A computer-readable medium (14) having stored thereon the computer program product of claim 18.