Method and device for improving robustness of photovoltaic prediction of multi-task space-time diagram fusion model

Abstract

The invention provides a method and device for improving robustness of photovoltaic prediction of a multi-task space-time diagram fusion model, and the method comprises the steps: data preprocessing: deleting abnormal values in photovoltaic power generation data and weather data, and carrying out the forward interpolation filling of missing values; a GAT-LSTM multi-task space-time diagram model is constructed; an L-CATBOOST cycle prediction model is constructed; performing multi-task standard deviation reciprocal method fusion: calculating historical prediction error standard deviation of the GAT-LSTM multi-task space-time diagram model and the L-CATBOOST cycle prediction model at each photovoltaic station; dynamically allocating the weight of the L-CATBOOST cycle prediction model according to the reciprocal of the standard deviation; and carrying out weighted fusion on prediction results of the GAT-LSTM multi-task space-time diagram model and the L-CATBOOST loop prediction model, and outputting a final prediction value. According to the invention, the standard deviation of the error of each target can be effectively reduced, so that the fluctuation of the error of each target is reduced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence algorithms, and specifically to a method and device for improving the robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model. Background Art

[0002] As the proportion of renewable energy in new power systems increases, the randomness and volatility of photovoltaic power generation poses challenges to power system stability. Furthermore, difficulties in integrating renewable energy into the grid have led to widespread curtailment of wind and solar power. This increased uncertainty in user-side electricity demand has posed new requirements for the efficient use of renewable energy, making photovoltaic power generation forecasting particularly necessary.

[0003] While the accuracy and error of time series forecasting models typically receive considerable attention, the stability of model predictions is also crucial in engineering applications. This refers to the volatility of the error in the model's predictions over time, quantified by the standard deviation (the standard deviation of the model error). Existing time series forecasting models focus on accuracy but overlook the impact of error volatility. Furthermore, multi-task forecasting models (which simultaneously predict multiple dependent variables Y) are more complex than single-task forecasting models (which predict only one dependent variable Y at a time), and their stability warrants even greater attention. Summary of the Invention

[0004] In order to address the above shortcomings of the multi-task prediction model, the present invention proposes a method and device for improving the robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model.

[0005] A method for improving the robustness of photovoltaic forecasting using a multi-task spatiotemporal graph fusion model includes the following steps:

[0006] (1) Data preprocessing: Delete outliers in photovoltaic power generation data and weather data, and fill missing values by forward interpolation;

[0007] (2) Constructing a GAT-LSTM multi-task spatiotemporal graph model:

[0008] A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold.

[0009] The graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the predicted power generation value of each photovoltaic site in the future preset period.

[0010] (3) Constructing L-CATBOOST cycle prediction model:

[0011] Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations;

[0012] A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results;

[0013] (4) Multi-task standard deviation inverse fusion:

[0014] Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site;

[0015] The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight.

[0016] The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

[0017] Furthermore, the data preprocessing in step (1) specifically includes:

[0018] Delete abnormal records with values of 999999 in the visibility, humidity, temperature, and pressure fields;

[0019] After replacing string or garbled code outliers with null values, forward interpolation is used to fill in the values.

[0020] Furthermore, the parameters of the bidirectional LSTM network in step (2) are set as:

[0021] The number of hidden layer units is 64, the input sequence length is 24 hours, and the activation function is ReLU;

[0022] The loss function is mean square error (MSE), the optimizer is Adam, and the learning rate is 0.001-0.005.

[0023] Furthermore, the construction of the spatiotemporal graph structure in step (2) satisfies the following conditions:

[0024] The calculation period of the Pearson correlation coefficient is the historical 24-hour data, and the preset threshold of the correlation coefficient is set to 0.5;

[0025] The attention score is calculated using a multi-head graph attention mechanism, with 4 to 8 attention heads and 8 to 16 output feature dimensions.

[0026] Furthermore, the training process of the L-CATBOOST cyclic prediction model in step (3) includes:

[0027] Train the CatBoost sub-model independently for each PV site and save them as model1 to model6;

[0028] In the prediction stage, the historical prediction value is used as one of the input features, and is updated in a rolling manner according to the time step to cyclically predict the power generation power in the next 24 hours.

[0029] Furthermore, the weight distribution formula in step (4) is:

[0030]

[0031] Among them, σ L，i and σ G，i are the standard deviations of the prediction errors of GAT-LSTM and L-CATBOOST at the i-th photovoltaic site, and satisfy W G，i +W L，i =1.

[0032] A device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model, comprising:

[0033] Data preprocessing module, used to remove outliers in photovoltaic power generation data and weather data, and to perform forward interpolation to fill missing values;

[0034] GAT-LSTM multi-task spatiotemporal graph model building module for:

[0035] A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold.

[0036] The graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the predicted power generation value of each photovoltaic site in the future preset period.

[0037] L-CATBOOST cycle prediction model building blocks for:

[0038] Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations;

[0039] A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results;

[0040] Multi-task inverse standard deviation fusion module, used for:

[0041] Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site;

[0042] The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight.

[0043] The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

[0044] Furthermore, the construction of the space-time graph structure satisfies the following conditions:

[0045] The calculation period of the Pearson correlation coefficient is the historical 24-hour data, and the preset threshold of the correlation coefficient is set to 0.5;

[0046] The attention score is calculated using a multi-head graph attention mechanism, with 4 to 8 attention heads and 8 to 16 output feature dimensions.

[0047] Furthermore, the training process of the L-CATBOOST cycle prediction model includes:

[0048] Train the CatBoost sub-model independently for each PV site and save them as model1 to model6;

[0049] In the prediction stage, the historical prediction value is used as one of the input features, and is updated in a rolling manner according to the time step to cyclically predict the power generation power in the next 24 hours.

[0050] Furthermore, the weight distribution formula is:

[0051]

[0052] Among them, σ L，i and σ G，i are the standard deviations of the prediction errors of GAT-LSTM and L-CATBOOST at the i-th photovoltaic site, and satisfy W G，i +W L，i =1.

[0053] The present invention proposes a fusion model aimed at reducing error volatility. Based on a spatiotemporal graph model and a time series model, this fusion model uses the inverse multi-task standard deviation method to weight the GAT-LSTM multi-task model and the L-CATBOOST cyclic prediction model (to distinguish it from conventional CATBOOST, L-CATBOOST is used here to represent the CATBOOST cyclic prediction model). This can effectively reduce the standard deviation of the error of each target, thereby reducing the volatility of the error of each target, and significantly enhancing the stability of the fused model. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a method for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to the present invention.

[0055] Figure 23. It is a schematic diagram comparing the robustness of the fusion model of an embodiment of the present invention and the GAT-LSTM multi-task prediction model and the L-CATBOOST cyclic prediction model. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] See also Figure 1 The embodiment of the present invention provides a method for improving the robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model. Figure 1 The left side of the figure shows the multi-task prediction model GAT-LSTM, and the right side shows the L-CATBOOST cyclic prediction model. The multi-task standard deviation inverse method is used for model fusion. The method includes the following steps:

[0058] (1) Data preprocessing: Delete outliers in photovoltaic power generation data and weather data, and fill missing values by forward interpolation;

[0059] (2) Constructing a GAT-LSTM multi-task spatiotemporal graph model:

[0060] A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold.

[0061] A graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the power generation prediction value of each photovoltaic site in the future preset period.

[0062] (3) Constructing L-CATBOOST cycle prediction model:

[0063] Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations;

[0064] A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results;

[0065] (4) Multi-task standard deviation inverse fusion:

[0066] Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site;

[0067] The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight.

[0068] The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

[0069] The data preprocessing in step (1) specifically includes:

[0070] Delete abnormal records with values of 999999 in the visibility, humidity, temperature, and pressure fields;

[0071] After replacing string or garbled code outliers with null values, forward interpolation is used to fill in the gaps.

[0072] Among them, the parameters of the bidirectional LSTM network in step (2) are set as:

[0073] The number of hidden layer units is 64, the input sequence length is 24 hours, and the activation function is ReLU;

[0074] The loss function is mean square error (MSE), the optimizer is Adam, and the learning rate is 0.001-0.005.

[0075] The construction of the spatiotemporal graph structure in step (2) satisfies the following conditions:

[0076] The calculation period of the Pearson correlation coefficient is the historical 24-hour data, and the preset threshold of the correlation coefficient is set to 0.5;

[0077] The attention score is calculated using a multi-head graph attention mechanism, with 4 to 8 attention heads and 8 to 16 output feature dimensions.

[0078] The training process of the L-CATBOOST cyclic prediction model in step (3) includes:

[0079] Train the CatBoost sub-model independently for each PV site and save them as model1 to model6;

[0080] In the prediction stage, the historical prediction value is used as one of the input features, and is updated in a rolling manner according to the time step to cyclically predict the power generation power in the next 24 hours.

[0081] Among them, the weight distribution formula in step (4) is:

[0082]

[0083] Among them, σ L，i and σ G，i are the standard deviations of the prediction errors of GAT-LSTM and L-CATBOOST at the i-th photovoltaic site, and satisfy W G，i +W L，i =1.

[0084] The technical solution of the present invention is described in detail below with a specific embodiment:

[0085] 1.1 Obtaining Data

[0086] The dataset is sourced from the Data Center, and is divided into photovoltaic power generation data and weather data. The period is from July 1, 2023, to January 31, 2024, in W City, with an hourly granularity. Weather data is obtained from the W City Meteorological Bureau and imported into the Data Center. The specific dimensions are as follows:

[0087] (1) Photovoltaic data: date, six photovoltaic power stations at a certain location in City W (generated power P1 to P6);

[0088] (2) Weather data: date, visibility, humidity, temperature, station pressure, and rainfall.

[0089] 1.2 Data Preprocessing

[0090] 1. Delete the abnormal value 999999 of visibility, which indicates that 999999 is mistakenly used for padding. The normal maximum value usually does not exceed 1000.

[0091] 2. Delete the abnormal value 999999 for humidity. This value indicates that 999999 was mistakenly filled in here. The normal maximum value usually does not exceed 100.

[0092] 3. Delete the abnormal temperature value 999999, which indicates that 999999 was mistakenly filled in here. The normal maximum value usually does not exceed 100;

[0093] 4. Delete the abnormal value 999999 of the air pressure. This value indicates that 999999 was mistakenly filled in here. The normal maximum value usually does not exceed 1050;

[0094] 5. For other abnormal values, such as strings, garbled characters, etc., replace them with null values and then use interpolation (forward interpolation) to fill them.

[0095] 1.3 Setting up the model

[0096] (1) Establishing a GAT-LSTM multi-task spatiotemporal graph model

[0097] The spatiotemporal graph network GAT-LSTM consists of two components: the GAT and the LSTM. The first step is to construct the graph structure required by the GAT algorithm, including the graph nodes and edges. In this paper, since there are six PV sites, each site is set as a graph node, for a total of six graph nodes. Pearson correlations are then performed on each of the six PV sites, with correlations greater than 0.5 considered as edges between any two graph nodes. If the correlation between two nodes is less than 0.5, there is no edge connecting the two graph nodes. Since the six PV sites in City W are geographically close, the power generation between them is closely correlated, with Pearson correlations greater than 0.5, meaning each graph node has five edges. Furthermore, each node has an attention score that measures the degree of connection with its neighboring nodes. These attention scores are learned and dynamically adjusted based on the characteristics of the nodes.

[0098] The data is then passed to a bidirectional long short-term memory (LSTM) network. The features extracted by the graph attention network at different times are fed into the bidirectional LSTM network in a time series format to learn temporal features. Using ReLU as the activation function, a batch size of 30, a 24-hour history period, and six fully connected layers as synchronous outputs, the network performs hourly single-step rolling predictions of the power generation of six photovoltaic sites.

[0099] (2) Establishing L-CATBOOST cycle prediction model

[0100] CatBoost is a machine learning algorithm based on the gradient boosting framework. Compared to other gradient boosting algorithms, CatBoost offers superior performance and efficiency. First, a CatBoost prediction model is established. The optuna algorithm library is then used to fit and optimize CatBoost's hyperparameters. After obtaining the optimal hyperparameters, a recurrent prediction model, L-CATBOOST, is then established. The entire CATBOOST preprocessing and training process is encapsulated and called using a for function loop. Each model trained using historical data is saved, resulting in models 1 through 6. Finally, predictions are looped over the trained models.

[0101] 1.4 Multi-task inverse standard deviation fusion

[0102] This step of model fusion is the key to the present invention. The multi-task standard deviation inverse method is used for fusion. The prediction errors of the previous two models at each photovoltaic site are calculated based on historical data. Then, the standard deviation of the error of each photovoltaic site of each model is calculated and weighted accordingly. The larger the standard deviation, the lower the weight, and the smaller the standard deviation, the higher the weight. The same photovoltaic site is a group of weights, and the weights of the two model results here add up to 1. Since there are 6 photovoltaic sites, there are a total of 6 groups of weight values. Finally, in the formal prediction stage, GAT-LSTM and L-CATBOOST are used for prediction respectively, and the final result is obtained after weighting according to the calculated weight values. The specific process is as follows:

[0103] The error e is calculated between the predicted Yp value for each task (i.e., each site i) and the true value Yt according to the historical data model:

[0104]

[0105] Next, we calculate the standard deviation of the error, where u is the mean of the error, as follows:

[0106]

[0107] Calculate the weights of the two models based on the standard deviation of the error:

[0108]

[0109] After fusion, the predicted value of the fusion model is obtained:

[0110] F p,i =W G,i *f G,i +W L,i *f L,i (5)

[0111] Finally, the robustness of the fusion model is calculated (the standard deviation of the overall average error):

[0112]

[0113] 1.5 Results Analysis

[0114] The 24-hour photovoltaic power generation of six photovoltaic sites on 2024-02-01 was predicted, and the mape error for each hour was calculated. Finally, the standard deviation of the prediction error for each site during this 24-hour period was calculated to measure the robustness of the model. A larger standard deviation indicates lower stability, and vice versa. The standard deviation of the prediction error of each model is shown in Table 1 below:

[0115] Table 1

[0116]

[0117] like Figure 2 As shown in the radar chart, the values 1 to 6 represent the standard deviation of the model error, and P1 to P6 represent six photovoltaic sites. The standard deviation of the error of the fusion model at each site is small, and the area of the enclosed polygon is also the smallest, showing relatively optimal robustness.

[0118] An embodiment of the present invention further provides a device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model, including:

[0119] Data preprocessing module, used to remove outliers in photovoltaic power generation data and weather data, and to perform forward interpolation to fill missing values;

[0120] GAT-LSTM multi-task spatiotemporal graph model building module for:

[0121] A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold.

[0122] The graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the predicted power generation value of each photovoltaic site in the future preset period.

[0123] L-CATBOOST cycle prediction model building blocks for:

[0124] Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations;

[0125] A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results;

[0126] Multi-task inverse standard deviation fusion module, used for:

[0127] Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site;

[0128] The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight.

[0129] The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for improving the robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model, characterized by: The steps include: (1) Data preprocessing: Delete outliers in photovoltaic power generation data and weather data, and fill missing values by forward interpolation; (2) Constructing a GAT-LSTM multi-task spatiotemporal graph model: A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold. The graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the predicted power generation value of each photovoltaic site in the future preset period. (3) Constructing L-CATBOOST cycle prediction model: Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations; A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results; (4) Multi-task standard deviation inverse fusion: Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site; The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight. The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

2. The method for improving robustness of photovoltaic forecasting using a multi-task spatiotemporal graph fusion model according to claim 1, characterized in that: The data preprocessing in step (1) specifically includes: Delete abnormal records with values of 999999 in the visibility, humidity, temperature, and pressure fields; After replacing string or garbled code outliers with null values, forward interpolation is used to fill in the gaps.

3. The method for improving robustness of photovoltaic forecasting using a multi-task spatiotemporal graph fusion model according to claim 1, characterized in that: The parameters of the bidirectional LSTM network in step (2) are set as: The number of hidden layer units is 64, the input sequence length is 24 hours, and the activation function is ReLU; The loss function is mean square error (MSE), the optimizer is Adam, and the learning rate is 0.001-0.

005.

4. The method for improving robustness of photovoltaic forecasting using a multi-task spatiotemporal graph fusion model according to claim 1, characterized in that: The construction of the spatiotemporal graph structure in step (2) satisfies the following conditions: The calculation period of the Pearson correlation coefficient is the historical 24-hour data, and the preset threshold of the correlation coefficient is set to 0.5; The attention score is calculated using a multi-head graph attention mechanism, with 4 to 8 attention heads and 8 to 16 output feature dimensions.

5. The method for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to claim 1, characterized in that: The training process of the L-CATBOOST cyclic prediction model in step (3) includes: Train the CatBoost sub-model independently for each PV site and save them as model1 to model6; In the prediction stage, the historical prediction value is used as one of the input features, and is updated in a rolling manner according to the time step to cyclically predict the power generation power in the next 24 hours.

6. The method for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to claim 1, characterized in that: The weight distribution formula in step (4) is: Among them, σ L，i and σ G，i are the standard deviations of the prediction errors of GAT-LSTM and L-CATBOOST at the i-th photovoltaic site, and satisfy W G，i +W L，i =1.

7. A device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model, characterized in that: include: Data preprocessing module, used to delete outliers in photovoltaic power generation data and weather data, and to perform forward interpolation to fill missing values; GAT-LSTM multi-task spatiotemporal graph model building module for: A spatiotemporal graph structure is constructed based on the Pearson correlation coefficient between PV sites. The spatiotemporal graph structure includes nodes and edges, where nodes are PV sites and edges are connections between two PV sites whose Pearson correlation coefficient is greater than a preset threshold. The graph attention network (GAT) is used to extract node spatial features, and the spatiotemporal features are input into a bidirectional LSTM network to output the predicted power generation value of each photovoltaic site in the future preset period. L-CATBOOST cycle prediction model building blocks for: Use the Optuna algorithm to optimize the hyperparameters of the CatBoost model, including tree depth, learning rate, and number of iterations; A sub-model is trained independently for each PV site, and rolling prediction is achieved by cyclically calling historical prediction results; Multi-task inverse standard deviation fusion module, used for: Calculate the standard deviation of historical prediction errors of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model at each PV site; The weight of the L-CATBOOST cyclic prediction model is dynamically assigned based on the inverse of the standard deviation. The larger the standard deviation, the lower the model weight. The prediction results of the GAT-LSTM multi-task spatiotemporal graph model and the L-CATBOOST cyclic prediction model are weighted and fused to output the final prediction value.

8. The device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to claim 7, characterized in that: The construction of the space-time graph structure meets the following conditions: The calculation period of the Pearson correlation coefficient is the historical 24-hour data, and the preset threshold of the correlation coefficient is set to 0.5; The attention score is calculated using a multi-head graph attention mechanism, with 4 to 8 attention heads and 8 to 16 output feature dimensions.

9. The device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to claim 7, characterized in that: The training process of the L-CATBOOST cycle prediction model includes: Train the CatBoost sub-model independently for each PV site and save them as model1 to model6; In the prediction stage, the historical prediction value is used as one of the input features, and is updated in a rolling manner according to the time step to cyclically predict the power generation power in the next 24 hours.

10. The device for improving robustness of photovoltaic prediction using a multi-task spatiotemporal graph fusion model according to claim 7, characterized in that: The weight distribution formula is: Among them, σ L，i and σ G，i are the standard deviations of the prediction errors of GAT-LSTM and L-CATBOOST at the i-th photovoltaic site, and satisfy W G，i +W L，i =1.

Images