Total radiation forecasting model training method, total radiation forecasting method and medium

Abstract

The invention belongs to the field of photovoltaic prediction, and particularly relates to a total radiation prediction model training method, a prediction method and a medium, and the training method comprises the steps: forming an observation data subset through total radiation live observation values; obtaining original forecast data of multiple numerical modes, and forming forecast data subsets after spatial downscaling; determining a first characteristic factor corresponding to each numerical mode based on the two subsets; forming a training data set by the two subsets, performing prediction based on the first characteristic factor and the training data set by using a gradient lifting decision tree to obtain initial prediction results in each numerical mode, and forming a prediction data set; determining a second feature factor based on the prediction data set and the observation data subset by using a deep learning model, and performing integration based on the second feature factor and the prediction data set to obtain a total radiation prediction result; and adjusting and optimizing parameters of the deep learning model to obtain a total radiation forecasting model. According to the method, the total radiation prediction is performed by integrating spatial downscaling, initial prediction and multi-mode integration, and the prediction accuracy can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of photovoltaic prediction, and in particular relates to a total radiation prediction model training method, a total radiation prediction method and a medium. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The accuracy of photovoltaic power generation prediction is largely determined by the accuracy of the forecast of total radiation meteorological elements. At present, the short- and medium-term total radiation meteorological forecast is highly dependent on the numerical weather forecast model, but it is limited by the boundary conditions, initial field, spatial resolution and other factors of the numerical forecast model. In particular, since photovoltaic power stations are mostly built in areas far away from national meteorological stations, the effect of numerical model data assimilation is not good, which makes the total radiation forecast accuracy directly output by the numerical forecast model cannot fully meet the business requirements of photovoltaic power stations. In order to further improve the accuracy of total radiation forecast, it is necessary to post-process the numerical model forecast.

[0004] For this purpose, machine learning models or deep learning models are proposed to post-process the numerical model forecasts. Although the accuracy of the total radiation forecast can be improved to a certain extent, the station forecasts in the existing methods are mostly direct forecasts from grids of different scales (i.e., different spatial resolutions) to the station. There is a lack of the process of refining the forecasts of different scales and then sending them to the station. As a result, some forecast sources with coarse spatial resolution cannot well express the small-scale forecasts of the station. In addition, the existing models usually do not consider the meteorological physical mechanism and the atmospheric circulation configuration as a whole when selecting characteristic factors, resulting in low forecast accuracy, which makes it difficult to meet the grid's assessment requirements for the forecast accuracy of photovoltaic power stations. Summary of the invention

[0005] In order to overcome the shortcomings of the background technology, the present invention provides a total radiation forecast model training method, a total radiation forecast method and a medium. The training method adopts spatial downscaling to unify the spatial resolution, constructs a first characteristic factor and a second characteristic factor for prediction, and uses a deep learning model for multi-mode integration to obtain a total radiation prediction result, thereby improving the accuracy of the total radiation forecast model prediction.

[0006] The technical solution adopted by the present invention is: a total radiation forecast model training method, the method comprising:

[0007] Obtain the actual total radiation observation value of the target point to form an observation data subset;

[0008] The original forecast data of multiple numerical models are obtained, the spatial resolutions of the original forecast data of different numerical models are uniformly downscaled to a preset resolution to form refined forecast data, and the refined forecast data of different numerical models are combined into a forecast data subset;

[0009] Determine a first characteristic factor corresponding to each numerical mode based on the forecast data subset and the observation data subset;

[0010] The forecast data subset and the observation data subset are combined into a training data set, and a gradient boosting decision tree algorithm is used to perform prediction based on the first characteristic factor and the training data set to obtain an initial prediction result of the gradient boosting decision tree algorithm under each numerical mode, and the initial prediction results are combined into a prediction data set;

[0011] Determine a second characteristic factor for multi-mode integration based on the prediction data set and the observation data subset using a deep learning model, and perform multi-mode integration based on the second characteristic factor and the prediction data set to obtain a total radiation prediction result;

[0012] A residual value is calculated based on the total radiation prediction result, parameters of the deep learning model are adjusted based on the residual value, and a total radiation forecast model is obtained when the residual value reaches a preset residual threshold.

[0013] Further,

[0014] The bilinear interpolation method is used to uniformly downscale the spatial resolution of the original forecast data from different numerical models to the preset resolution to form refined forecast data.

[0015] Further,

[0016] The determining of the first characteristic factor corresponding to each numerical mode based on the forecast data subset and the observation data subset includes:

[0017] Determine the label value based on the observed data subset; use the gradient boosting decision tree algorithm to sort the importance of forecast factors based on the forecast data subset and the label value, wherein the forecast factors are factors affecting radiation forecast; and determine the first characteristic factor corresponding to each numerical pattern based on the sorting result.

[0018] Further,

[0019] The method of using a deep learning model to determine a second characteristic factor for multi-modal integration based on the prediction data set and the observation data subset includes:

[0020] The deep learning model adopts a deep residual network, and uses the deep residual network to associate and perform feature fusion learning on the initial prediction results in the prediction data set and the actual total radiation observation values ​​in the observation data subset, screen out the most critical forecasting elements for total radiation forecast, and train the weight distribution of each forecasting element in combination with the attention mechanism, and determine the second characteristic factor for multi-mode integration based on the weight distribution. The forecasting element is the element that affects the radiation forecast.

[0021] Further,

[0022] The method of using a deep learning model to perform multi-mode integration based on the second characteristic factor and the prediction data set to obtain a total radiation prediction result includes:

[0023] The deep learning model adopts a deep residual network. According to the different forecast time effects of the initial prediction results under different numerical modes, a backbone network and a branch network of the deep residual network are designed. The second feature factor is then input into the deep residual network. The backbone network and the branch network are used to extract features respectively, and the weight distribution of each feature is trained in combination with the attention mechanism. Finally, the features are spliced ​​and input into a decoder to obtain the total radiation prediction result.

[0024] Further,

[0025] Based on the second characteristic factor, the constructed deep residual network is trained and weight optimized, and the Bayesian optimization algorithm is used to optimize the parameters of the deep residual network.

[0026] Further,

[0027] Both the first characteristic factor and the second characteristic factor include one or more of meteorological characteristic factors and topographic characteristic factors; wherein the meteorological characteristic factors include one or more of downward shortwave radiation, direct radiation, scattered radiation, temperature, humidity, wind speed, wind direction, albedo, rainfall, snowfall, and cloud cover; and the topographic characteristic factors include one or more of longitude, latitude, and altitude.

[0028] Further,

[0029] After obtaining the initial prediction results of the gradient boosting decision tree algorithm in each numerical mode, the nearest point method is used to select the initial prediction results of a grid point closest to the longitude and latitude of the target point to form the prediction data set.

[0030] The present invention also provides a total radiation forecasting method, comprising: acquiring forecast data and observation data; inputting the forecast data and the observation data into a total radiation forecasting model; wherein the total radiation forecasting model is obtained by the above-mentioned total radiation forecasting model training method; the total radiation forecasting model processes the forecast data and the observation data, and outputs a total radiation prediction result.

[0031] The present invention also provides a computer storage medium storing one or more programs, which, when executed, can implement the aforementioned total radiation forecast model training method, or the aforementioned total radiation forecast method.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The forecast data subset includes refined forecast data of multiple numerical models, which not only eliminates the defect of poor stability of single-model forecast, but also unifies the spatial resolution of forecasts of different numerical models through spatial downscaling, which can improve the radiation forecast effect;

[0034] 2. In the construction of the first characteristic factor, a variety of numerical modes are considered to make the prediction of the gradient boosting decision tree algorithm more accurate and enhance the interpretability of the weather process. In the construction of the second characteristic factor, a deep residual network is used to determine the second characteristic factor for multi-mode integration based on the prediction data set and the observation data subset, which can effectively remove some elements that are weakly associated with the total radiation forecast, thereby enhancing the generalization ability of the model;

[0035] 3. The deep residual network is used to perform multi-modal integration based on the second characteristic factor and the predicted data set to obtain the total radiation prediction results. At the same time, the attention mechanism is introduced to increase the interpretability of the model. The Bayesian optimization algorithm is also used to optimize the parameters of the deep residual network model, thereby improving the accuracy of the total radiation forecast model prediction.

[0036] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings.

[0037] The present invention will be further described below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0039] Figure 1 A schematic diagram of a flow chart of a total radiation forecast model training method according to an embodiment of the present invention;

[0040] Figure 2 The figure is a flow chart of a total radiation forecasting method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the implementation of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, this embodiment discloses a total radiation forecast model training method, the method comprising:

[0043] S1. Obtain the actual total radiation observation value of the target point to form an observation data subset;

[0044] In the embodiment of the present disclosure, the photovoltaic power station is taken as the target point, and the observation data subset includes the actual total radiation observation value of the target point. The actual total radiation observation value is the observation data of the total radiation of the photovoltaic power station.

[0045] In some embodiments, the observation data subset is obtained by the following steps: acquiring the total radiation observation data of the photovoltaic power station; removing the total radiation observation values ​​that exceed a preset radiation threshold in the total radiation observation data; and / or removing the continuous and unchanged total radiation observation values ​​that exceed a preset time length to obtain the observation data subset.

[0046] The radiation threshold can be set by the user according to the characteristics of different regions, and the embodiment of the present disclosure does not limit this. For example, the radiation threshold can be 1400W / m 2 .

[0047] For example, according to the "Solar Energy Resource Assessment Method GB / T 37526-2019" and the meteorological industry standard "Meteorological Observation Data Quality Control Ground" (QX / T 118-2020), the total radiation observation data of the photovoltaic power station is quality controlled, and the total radiation observation data with observation values ​​exceeding 1400W / m2 and the total radiation observation values ​​that remain unchanged for more than 5 hours are treated as missing. The singular values ​​and erroneous values ​​in the total radiation observation data are removed, and the seasonal and diurnal changes of the observation data are further eliminated through data analysis, and finally a structured observation data subset is obtained.

[0048] In some embodiments, the observation data in the observation data subset is stored in units of months. For subsequent model training, the observation data of the first 24 days of each month can be used as a training set, and the observation data after 25 days can be used as a test set.

[0049] S2. obtaining original forecast data of multiple numerical models, uniformly downscaling the spatial resolution of the original forecast data of different numerical models to a preset resolution to form refined forecast data, and composing the refined forecast data of different numerical models into a forecast data subset;

[0050] The present embodiment includes the original forecast data of multiple numerical modes, which can be obtained by the neighboring point method. In some embodiments, the original forecast data of multiple numerical modes include the original forecast data of at least two data sources in the China Meteorological Administration Wind Energy and Solar Energy Prediction System (CMA-WSP), the China Meteorological Administration Regional Mesoscale Numerical Prediction System (CMA-MESO), the China Meteorological Administration Guangdong Rapid Update Assimilation Numerical Prediction System (CMA-GD), and the Solar Weather Research and Forecast Model-Solar (WRF-SOLAR). In the specific implementation, the original forecast data of the multi-source numerical mode is also preprocessed, mainly including the unified processing of data format, unit, etc.

[0051] The original forecast data of various numerical models come from different original data sources and are based on grid scales. Different numerical model forecasts often have different spatial resolutions, which are generally greater than 1 km, such as 9 km, 3 km, etc. In order to further improve the total radiation forecast effect, the present invention uses spatial downscaling to unify the spatial resolution.

[0052] In this embodiment, the preset spatial resolution is 1KM, that is, the spatial resolutions of the original forecast data of different numerical modes are uniformly downscaled to 1KM to form refined forecast data.

[0053] In some embodiments, the spatial resolution of the original forecast data of different numerical modes is uniformly downscaled to 1KM using bilinear interpolation to form refined forecast data. Spatial downscaling based on bilinear interpolation can achieve a uniform forecast spatial resolution of 1KM in the forecast area centered on the photovoltaic power station (i.e., the target point), which can improve the forecast accuracy to a certain extent and solve the problems existing in the prior art.

[0054] This embodiment takes the surrounding influence into consideration, and for the original forecast data of any numerical mode, not only the original forecast data of the target point is extracted, but also the original forecast data of four grid points adjacent to the target point, above, below, left and right, are extracted.

[0055] For example, taking the photovoltaic power station as the target point, the adjacent grid point method is used to extract the original forecast data of the four grid points adjacent to the target point, that is, the original forecast data of 5 grid points are extracted for each photovoltaic power station, and the original forecast data of the 5 grid points are spatially downscaled to 1KM to form refined forecast data, and then the refined forecast data constitutes a forecast data subset.

[0056] In some embodiments, the refined forecast data in the forecast data subset is stored in units of months. For subsequent model training, the refined forecast data for the first 24 days of each month can be used as a training set, and the refined forecast data after 25 days can be used as a test set.

[0057] S3. Determine a first characteristic factor corresponding to each numerical mode based on the forecast data subset and the observed data subset;

[0058] This embodiment determines the first characteristic factor based on the forecast data subset and the observation data subset. Specifically, the label value is determined based on the observation data subset; the forecast elements are ranked in importance based on the forecast data subset and the label value using the gradient boosting decision tree (GBDT) algorithm, and the forecast elements are elements that affect the radiation forecast; the first characteristic factor corresponding to each numerical pattern is determined based on the sorting result. Since the actual observation value of total radiation is a real measured value, it can be used as the label value of the training model. This embodiment uses the actual observation value of total radiation in the observation data subset as the label value. In this embodiment, the data in the forecast data subset can be understood as forecast elements.

[0059] For example, the refined forecast data of multiple numerical modes are used as feature vectors, and the actual total radiation observation value is input into the GBDT model as the label value. The feature factors output by the GBDT model are sorted according to importance, and the most important feature factor is selected as the first feature factor. The gradient boosting decision tree (GBDT) algorithm belongs to the prior art and will not be described in detail here.

[0060] It should be noted that for the data sources of the four numerical models CMA-WSP, CMA-MESO, CMA-GD, and WRF--SOLAR, since the forecast effects of each data source are different, the first characteristic factor selected in each data source may also be different. For example, albedo is selected as the first characteristic factor in the CMA-WS numerical model, but albedo may not be selected in the CMA-GD numerical model.

[0061] The disclosed embodiments are based on the principle of characterizing the physical process of meteorology, taking into account the forecasting factors affecting the radiation circulation and at least one of the forecasting factors of geographical factors. Among them, the forecasting factors affecting the radiation circulation include but are not limited to long-wave and short-wave radiation, direct radiation, scattered radiation, temperature, humidity, wind speed, wind direction, albedo, rainfall, snowfall, cloud cover and other factors. The forecasting factors of geographical factors include but are not limited to longitude, latitude, and altitude.

[0062] In this embodiment, the above-mentioned first characteristic factor and the second characteristic factor described later both include one or more of meteorological characteristic factors and topographic characteristic factors; wherein the meteorological characteristic factors include one or more of downward shortwave radiation, direct radiation, scattered radiation, temperature, humidity, wind speed, wind direction, albedo, rainfall, snowfall, and cloud cover; and the topographic characteristic factors include one or more of longitude, latitude, and altitude.

[0063] S4. Combining the forecast data subset and the observation data subset into a training data set, using the gradient boosting decision tree algorithm to perform prediction based on the first feature factor and the training data set, obtaining the initial prediction result of the gradient boosting decision tree algorithm under each numerical mode, and combining the initial prediction results into a prediction data set;

[0064] This embodiment uses the gradient boosting decision tree (LightGBM) algorithm to make predictions based on the first characteristic factor and the training data set. The LightGBM model is an improved gradient boosting algorithm. The histogram-based segmentation algorithm replaces the pre-sort traversal algorithm, which has a significant improvement in training speed and can effectively prevent overfitting. It is more suitable for training massive high-dimensional data. During the prediction process, the LightGBM model adaptively constructs multiple decision trees according to the characteristics of the data for prediction to obtain initial prediction results. Each construction of a decision tree can make up for the error of the previous LightGBM model. In this embodiment, the forecast data subset and the observation data subset are both the initial data used for modeling, and the first characteristic factor is the data selected based on the forecast data subset and the observation data subset for input into the LightGBM model.

[0065] In some embodiments, 80% of the data is randomly selected as a training set for LightGBM model training and parameter tuning, and 20% of the data is used for initial prediction result verification and optimization, where the optimized parameters include the number of decision trees, the maximum depth of the decision tree, the number of iterations, etc. The algorithm is implemented based on the scikit-learn-Ran-domfrest data package in python. The spatial resolution of forecasts from different numerical models is different, generally greater than 1KM. Using the above-mentioned bilinear interpolation method and LightGBM model, forecasts from different numerical models are uniformly downscaled to 1KM before prediction, which not only improves the downscaling effect, but also helps to improve the prediction effect.

[0066] In some embodiments, after obtaining the initial prediction results of the gradient boosting decision tree algorithm in each numerical mode, the nearest point method is used to select the initial prediction results of a grid point closest to the longitude and latitude of the target point to form the prediction data set. In this embodiment, the photovoltaic power station is used as the target point, and the nearest point method is used to select the initial prediction results in different numerical modes, that is, the initial prediction results of a grid point closest to the longitude and latitude of the photovoltaic power station are selected as the prediction data set for integrated model training and testing.

[0067] S5. Determine a second characteristic factor for multi-mode integration based on the prediction data set and the observation data subset using a deep learning model, and perform multi-mode integration based on the second characteristic factor and the prediction data set to obtain a total radiation prediction result;

[0068] The deep learning model of this embodiment adopts a deep residual network (ResNet), which first selects the second characteristic factor for prediction input, and then uses the second characteristic factor to make the final integrated prediction.

[0069] In some embodiments, a deep residual network is used to associate and perform feature fusion learning on the initial prediction results in the prediction data set and the actual observation values ​​of total radiation in the observation data subset, screen out the most critical forecasting elements for total radiation forecast, and train the weight distribution of each forecasting element in combination with the attention mechanism, and determine the second characteristic factor for multi-mode integration based on the weight distribution, and the forecasting element is the element that affects the radiation forecast.

[0070] The use of a deep residual network can well perform feature analysis on the initial prediction results of different data sources and the actual total radiation observation values, and then select the forecasting elements that have the greatest impact on the total radiation forecast of the photovoltaic power station. This embodiment takes into account the impact of terrain on the climate element field, and can screen out key meteorological elements (i.e., meteorological characteristic factors) and high-resolution terrain (i.e., topographic characteristic factors) within the time window as features (i.e., the most critical forecasting elements), and combine the attention mechanism to train the weight distribution of each forecasting element, and finally determine the second feature factor for multi-mode integration based on the weight distribution. The attention mechanism is essentially a weighted mechanism, which realizes the mining of important features in the original data and strengthens its influence by weighted distribution of the output results of the deep residual network.

[0071] Before establishing a total radiation forecast model, it is necessary to extract key elements (i.e., the second characteristic factor) and remove some elements that are weakly related to the total radiation forecast. This can enhance the generalization ability of the model and reduce the training pressure of the model. The extracted key elements include meteorological elements and high-resolution terrain.

[0072] This embodiment integrates multi-source forecast data (i.e., initial forecast results under different numerical modes) based on deep learning. The deep residual network (ResNet) model is used to combine the spatial information of the forecast field around the site (i.e., the target point) and the time information of the model forecast, as well as the meteorological element information, to perform multi-mode integration of the total radiation elements of the site model forecast, and at the same time introduce an attention mechanism to increase the interpretability of the model.

[0073] In some embodiments, for different forecast time limits of the initial prediction results under different numerical modes, a backbone network and a branch network of a deep residual network are designed, and then the second feature factor is input into the deep residual network, and features are extracted using the backbone network and the branch network respectively, and the weight distribution of each feature is trained in combination with the attention mechanism, and finally the features are spliced ​​and input into the decoder to obtain the total radiation prediction result. For different forecast time limits of the initial prediction results of different data sources, a backbone network and a branch network are designed to extract features respectively, and spliced ​​and input into the decoder to obtain an integrated result. The backbone network and the branch network of the deep residual network of this embodiment are determined according to the forecast time limits of different numerical forecast modes. The numerical forecast modes with the same and longer forecast time limits can be used as the backbone network, and the numerical forecast modes with shorter forecast time limits can be used as the branch network. The above-mentioned decoder adopts an upsampling convolution decoder.

[0074] In some embodiments, the constructed deep residual network is trained and weight optimized based on the second characteristic factor, and the Bayesian optimization algorithm is used to optimize the parameters of the deep residual network. In this embodiment, a training set and a validation set are constructed based on the second characteristic factor, 80% of the data is used for the training set, and 20% of the data is used for the validation set, and the constructed model is trained and weight optimized using the training set and the validation set. Due to the high training cost of the hyperparameterization problem of deep learning, this embodiment uses the Bayesian optimization algorithm to optimize the parameters of the deep residual network model, first determine the model parameters according to the Bayesian optimization algorithm, and then train the model. Among them, the model parameters refer to the hyperparameters used in the integrated model, such as network structure parameters, convolutional layer parameters, etc. The Bayesian optimization algorithm uses a probabilistic proxy model to fit the true target function, and actively selects a point with a better fitting effect for the next forecast based on the fitting results. The Bayesian optimization algorithm mainly includes two key parts: a probabilistic proxy model and an acquisition function. Among them, the probabilistic proxy model uses a tree Parzen estimator (treeParzen estimator, TPE), and the acquisition function uses a Gaussian mixture model.

[0075] S6. Calculate a residual value based on the total radiation prediction result, adjust the parameters of the deep learning model based on the residual value, and obtain a total radiation forecast model when the residual value reaches a preset residual threshold.

[0076] In some embodiments, the total radiation prediction result is tested, and the test indicators include: mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R).

[0077] In the above-mentioned total radiation forecast model training method, since the training data includes a forecast data subset and an actual data subset, and the forecast data subset includes refined forecast data of multiple numerical modes, it not only eliminates the defect of poor stability of single-mode forecast, but also unifies the spatial resolution of forecasts of different numerical modes through spatial downscaling, which can improve the radiation forecast effect. In the construction of the first characteristic factor, multiple numerical modes are considered, which makes the prediction of the LightGBM model more accurate and can also enhance the interpretability of the weather process. In the construction of the second characteristic factor, a deep residual network (ResNet) is used to determine the second characteristic factor for multi-mode integration based on the forecast data set and the observation data subset, which can effectively remove some elements with weak correlation with the total radiation forecast, thereby enhancing the generalization ability of the model, and using the deep residual network (ResNet) to perform multi-mode integration based on the second characteristic factor and the forecast data set to obtain the total radiation prediction result. At the same time, the attention mechanism is introduced to increase the interpretability of the model, and the Bayesian optimization algorithm is used to optimize the parameters of the deep residual network model, thereby improving the accuracy of the prediction of the total radiation forecast model.

[0078] like Figure 2 As shown, the present invention also provides a total radiation forecasting method, comprising:

[0079] S1. Obtain forecast data and observation data;

[0080] S2. inputting the forecast data and the observation data into a total radiation forecast model; wherein the total radiation forecast model is obtained by the above-mentioned total radiation forecast model training method;

[0081] S3. The total radiation forecast model processes the forecast data and the observation data, and outputs a total radiation forecast result.

[0082] The total radiation forecasting method provided by the embodiment of the present disclosure has a more accurate and more explainable total radiation forecasting result because the total radiation forecasting model is a model obtained by the total radiation forecasting model training method provided by the embodiment of the present disclosure.

[0083] Based on the same inventive concept, the present invention also provides a computer storage medium storing one or more programs, which, when executed, can implement the aforementioned total radiation forecast model training method, or implement the aforementioned total radiation forecast method.

[0084] According to the total radiation forecast model training method and the total radiation forecast method of the embodiment of the present disclosure, the terminal device can be executed by an electronic device such as a terminal device or a server, and the terminal device can be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (Personal Digital Assistant, PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a cloud server capable of cloud computing. The method can be implemented by a processor calling a computer-readable program instruction stored in a memory.

[0085] Parts not involved in this embodiment are the same as the prior art or can be implemented by using the prior art, and will not be further described here.

[0086] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A total radiation forecast model training method, characterized in that: The method comprises: Obtain the actual total radiation observation value of the target point to form an observation data subset; The original forecast data of multiple numerical models are obtained, the spatial resolutions of the original forecast data of different numerical models are uniformly downscaled to a preset resolution to form refined forecast data, and the refined forecast data of different numerical models are combined into a forecast data subset; Determine a first characteristic factor corresponding to each numerical mode based on the forecast data subset and the observation data subset; The forecast data subset and the observation data subset are combined into a training data set, and a gradient boosting decision tree algorithm is used to perform prediction based on the first characteristic factor and the training data set to obtain an initial prediction result of the gradient boosting decision tree algorithm under each numerical mode, and the initial prediction results are combined into a prediction data set; Determine a second characteristic factor for multi-mode integration based on the prediction data set and the observation data subset using a deep learning model, and perform multi-mode integration based on the second characteristic factor and the prediction data set to obtain a total radiation prediction result; A residual value is calculated based on the total radiation prediction result, parameters of the deep learning model are adjusted based on the residual value, and a total radiation forecast model is obtained when the residual value reaches a preset residual threshold.

2. A total radiation forecast model training method according to claim 1, characterized in that: The bilinear interpolation method is used to uniformly downscale the spatial resolution of the original forecast data from different numerical models to the preset resolution to form refined forecast data.

3. A total radiation forecast model training method according to claim 1, characterized in that: The determining of the first characteristic factor corresponding to each numerical mode based on the forecast data subset and the observation data subset includes: Determine the label value based on the observed data subset; use the gradient boosting decision tree algorithm to sort the importance of forecast factors based on the forecast data subset and the label value, wherein the forecast factors are factors affecting radiation forecast; and determine the first characteristic factor corresponding to each numerical pattern based on the sorting result.

4. A total radiation forecast model training method according to claim 1, characterized in that: The method of using a deep learning model to determine a second characteristic factor for multi-modal integration based on the prediction data set and the observation data subset includes: The deep learning model adopts a deep residual network, and uses the deep residual network to associate and perform feature fusion learning on the initial prediction results in the prediction data set and the actual total radiation observation values ​​in the observation data subset, screen out the most critical forecasting elements for total radiation forecast, and train the weight distribution of each forecasting element in combination with the attention mechanism, and determine the second characteristic factor for multi-mode integration based on the weight distribution. The forecasting element is the element that affects the radiation forecast.

5. A total radiation forecast model training method according to any one of claims 1 to 4, characterized in that: The method of using a deep learning model to perform multi-mode integration based on the second characteristic factor and the prediction data set to obtain a total radiation prediction result includes: The deep learning model adopts a deep residual network. According to the different forecast time effects of the initial prediction results under different numerical modes, a backbone network and a branch network of the deep residual network are designed. The second feature factor is then input into the deep residual network. The backbone network and the branch network are used to extract features respectively, and the weight distribution of each feature is trained in combination with the attention mechanism. Finally, the features are spliced ​​and input into a decoder to obtain the total radiation prediction result.

6. A total radiation forecast model training method according to claim 5, characterized in that: Based on the second characteristic factor, the constructed deep residual network is trained and weight optimized, and the Bayesian optimization algorithm is used to optimize the parameters of the deep residual network.

7. A global radiation forecast model training method according to claim 1, characterized in that: Both the first characteristic factor and the second characteristic factor include one or more of meteorological characteristic factors and topographic characteristic factors; wherein the meteorological characteristic factors include one or more of downward shortwave radiation, direct radiation, scattered radiation, temperature, humidity, wind speed, wind direction, albedo, rainfall, snowfall, and cloud cover; and the topographic characteristic factors include one or more of longitude, latitude, and altitude.

8. A total radiation forecast model training method according to claim 1, characterized in that: After obtaining the initial prediction results of the gradient boosting decision tree algorithm in each numerical mode, the nearest point method is used to select the initial prediction results of a grid point closest to the longitude and latitude of the target point to form the prediction data set.

9. A total radiation forecasting method, characterized in that: include: Obtain forecast data and observation data; input the forecast data and the observation data into a total radiation forecast model; wherein the total radiation forecast model is obtained by the total radiation forecast model training method according to any one of claims 1 to 8; the total radiation forecast model processes the forecast data and the observation data, and outputs a total radiation prediction result.

10. A computer storage medium storing one or more programs, characterized in that: When the one or more programs are executed, the total radiation forecast model training method described in any one of claims 1 to 8 can be implemented, or the total radiation forecast method described in claim 9 can be implemented.

Images