A Multi-Task Learning Method for Coupled Information Separation and Extraction in Multi-Site Photovoltaic Power Prediction

Abstract

The application discloses a multi-station photovoltaic power prediction coupling information separation and extraction multi-task learning method, comprising the following steps: step S1: according to the power information, meteorological information and geographical position information of photovoltaic power stations, power similarity, meteorological similarity and geographical position similarity are calculated respectively, the correlation between photovoltaic power stations is comprehensively judged through comprehensive similarity, and similar power station groups are identified; step S2: a coupling information separation and extraction multi-task learning prediction model is constructed; step 3: a dynamic update loss function of the coupling information separation and extraction multi-task learning prediction model is constructed; and step 4: specific parameters of the coupling information separation and extraction multi-task learning prediction model are trained and selected. Through the construction of the task expert network specially extracting the specific information of each power station and the shared network extracting the coupling information between multiple power stations, the specific information and the coupling information can be separated, and the accuracy and robustness of the prediction model are improved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation prediction, and more specifically, to a multi-task learning method for coupling information separation and extraction in multi-site photovoltaic power prediction. Background Technology

[0002] Photovoltaic (PV) power generation forecasting is crucial for optimizing energy management and grid stability. Accurate forecasts can enhance the integration of solar energy with the power system, reduce reliance on fossil fuels, and minimize operating costs. By forecasting PV power generation, power companies can better balance supply and demand and improve the utilization of renewable energy.

[0003] Most existing photovoltaic (PV) power generation forecasting methods only consider the forecast of a single PV power plant, ignoring the correlation between different PV power plants. Since different PV power plants in the same region share similar geographical conditions and weather patterns, there are spatiotemporal coupling characteristics between them. Appropriately considering this coupling relationship can help improve the accuracy of PV power generation forecasting. In the field of power forecasting, multi-task learning is considered an effective method for uncovering the coupling relationships between different objects. However, traditional multi-task learning models cannot distinguish between the unique information of different forecasting tasks and the coupling information between tasks, leading to mutual interference during the forecasting process and generally poor forecasting results. Summary of the Invention

[0004] To address the problem that traditional multi-task learning models cannot distinguish between the unique information of different prediction tasks and the coupling information between tasks, this invention provides a multi-task learning method for separating and extracting coupling information in multi-station photovoltaic power prediction. This method can simultaneously predict the output power of multiple related photovoltaic power plants. During the prediction process, the unique information of each photovoltaic power plant can be separated from the coupling information between multiple photovoltaic power plants to avoid interference between information, thereby improving the accuracy and robustness of the prediction model.

[0005] To achieve the above objectives, this invention provides a multi-task learning method for separating and extracting coupled information in multi-site photovoltaic power prediction, comprising the following steps:

[0006] Step S1: Calculate power similarity, meteorological similarity, and geographical location similarity based on the power information, meteorological information, and geographical location information of the photovoltaic power station. Use the comprehensive similarity to judge the correlation between photovoltaic power stations and identify similar power station groups.

[0007] Step S2: Construct a coupled information separation and extraction multi-task learning and prediction model, which consists of an input layer, a feature extraction network, a prediction network, a gating network, and an output layer.

[0008] The input layer inputs a predicted input feature set of a similar power plant group;

[0009] The feature extraction network comprises a task expert network extracting information unique to each power plant and a shared network extracting coupling information between multiple power plants; the task expert network and the shared network are composed of multiple fully connected layer neural networks.

[0010] The gating network selectively fuses the outputs of the task expert network and the shared network; the output of the gating network of the kth task in the jth layer feature extraction network can be expressed as:

[0011] D k,j (x)=w D k,j (D k,j-1 (x))S k,j (x) (5);

[0012] wherein k represents the task number; x is the input data; w D k,j is the weight function of task k in the jth layer feature extraction network, which is calculated by the Softmax function: w D k,j =Softmax(W k,j x), wherein W k,j is a parameter matrix;

[0013] The prediction network takes the final output of the gating network as its input feature, which includes both the unique information of each photovoltaic power plant and the coupling information between different power plants; the prediction network is composed of a long short-term memory neural network.

[0014] The output layer is the power prediction value of each power plant;

[0015] Step 3: Construct a dynamic update loss function of the coupling information separation and extraction multi-task learning prediction model;

[0016] Step 4: Train and select the specific parameters of the coupling information separation and extraction multi-task learning prediction model.

[0017] Further, the power similarity is calculated, and the Pearson correlation coefficient is used to calculate the power similarity between different photovoltaic power plants:

[0018]

[0019] wherein P i and Q i are the power data of two power plants, and are their mean values.

[0020] Further, the weather similarity is calculated, and a multi-dimensional vector similarity measure is used to calculate the similarity of weather data:

[0021]

[0022] Wherein, M i , N i are the weather data vectors of two power stations respectively.

[0023] Further, when calculating the geographical position similarity, the Haversine formula is used to calculate the geographical distance between photovoltaic power stations:

[0024]

[0025] Wherein, φ and λ are the latitude and longitude respectively, and r is the radius of the earth.

[0026] Further, the comprehensive similarity is calculated:

[0027] S=w1·S power +w2·S weather +w3·S location (4);

[0028] Wherein, S power , S location and S weather are the power, weather and geographical position similarities respectively, and w1, w2, w3 are their weights.

[0029] Further, in step S1, similar power station groups are identified, and specifically, the DBSCAN clustering algorithm is used to group the power stations.

[0030] Further, the prediction input feature set includes: photovoltaic power historical data, prediction time, irradiance, temperature, humidity, cloud cover, rainfall.

[0031] Further, the dynamic updating loss function of the coupling information separation extraction multi-task learning prediction model is:

[0032]

[0033] In the formula: L k loss is the loss function of the kth task; a k is the weight of the kth loss function. According to the change rate, the weight of the loss function is dynamically updated:

[0034] a k (t)=a k,0 ×γ k (7);

[0035] In the formula: ak (t) is the weight of the kth loss function at the tth iteration; a k,0 is the initial value of a k is the initial value of a k is the change rate.

[0036] Further, the Adam optimization algorithm is used to solve the model parameters, and the cross-validation method is used to optimize and select the learning rate, the number of expert networks, the number of feature extraction network layers, the network layer number and other hyperparameters.

[0037] The beneficial effects of the present application are as follows:

[0038] 1. The present application proposes a coupled information separation extraction multi-task learning prediction method for multi-station photovoltaic power prediction, which can separate task-specific information and coupled information by constructing task-specific expert networks for extracting information specific to each power station and shared networks for extracting coupled information between multiple power stations, avoiding interference between information, thereby improving the accuracy and robustness of the prediction model.

[0039] 2. To solve the problem of difficult determination of different prediction task weights, the present application proposes a loss function dynamically updated according to a certain change rate, which helps to find the most suitable task weight and further improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description.

[0041] Figure 1 is a flow chart of a multi-station photovoltaic power prediction coupled information separation extraction multi-task learning method according to an embodiment of the present application;

[0042] Figure 2 is a coupled information separation extraction multi-task learning framework diagram proposed by an embodiment of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0044] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0045] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0046] As shown in Figure 1 , the embodiment of the present application provides a multi-station photovoltaic power prediction coupling information separation extraction multi-task learning method, comprising the following steps:

[0047] Step S1: According to the power information, meteorological information and geographical position information of the photovoltaic power station, the power similarity, the meteorological similarity and the geographical position similarity are calculated respectively, the correlation between the photovoltaic power stations is comprehensively judged through the comprehensive similarity, and the similar power station group is identified;

[0048] Specifically, when calculating the power similarity, the Pearson correlation coefficient is used to calculate the power similarity between different photovoltaic power stations:

[0049]

[0050] Wherein, P i and Q i are the power data of two power stations, and are their mean values.

[0051] When calculating the meteorological similarity, the multidimensional vector similarity measure (such as Euclidean distance, cosine similarity) is used to calculate the similarity of the meteorological data:

[0052]

[0053] Wherein, M i , N i are the meteorological data vectors of two power stations.

[0054] When calculating geographical similarity, the Haversine formula is used to calculate the geographical distance between photovoltaic power plants:

[0055]

[0056] Where φ and λ are latitude and longitude, respectively, and r is the Earth's radius.

[0057] Calculate the overall similarity:

[0058] S = w1·S power +w2·S weather +w3·S location (4);

[0059] Among them, S power S location and S weather These are the similarities in power, weather, and geographical location, respectively, with w1, w2, and w3 being their weights.

[0060] The DBSCAN clustering algorithm was used to group power plants and identify similar power plant groups.

[0061] Step S2: Construct a coupled information separation and extraction multi-task learning and prediction model. This model consists of an input layer, a feature extraction network, a prediction network, a gating network, and an output layer, as shown in the attached diagram. Figure 2 As shown.

[0062] The input layer takes into account the predicted input feature set of similar power plant groups, including: historical photovoltaic power data, prediction time, irradiance, temperature, humidity, cloud cover, rainfall, etc.

[0063] The feature extraction network includes a task expert network for extracting information specific to each power station and a shared network for extracting coupling information between multiple power stations. Both the task expert network and the shared network consist of multiple fully connected layer neural networks.

[0064] The gating network selectively fuses the outputs of the task expert network and the sharing network. The gating network output for the k-th task in the j-th layer feature extraction network can be expressed as:

[0065] D k,j (x)=w D k,j (D k,j-1 (x))S k,j (x) (5);

[0066] In the formula: k represents the task number; x is the input data; w D k,jThe weight function for task k in the j-th layer feature extraction network is calculated using the Softmax function: w D k,j =Softmax(W k,j x), where W k,j It is a parameter matrix.

[0067] The prediction network uses the final output of the gating network as its input features, which include both the unique information of each photovoltaic power station and the coupling information between different power stations. The prediction network is composed of a long short-term memory neural network.

[0068] The output layer contains the power prediction values ​​for each power station.

[0069] Step 3: Construct a dynamic update loss function for the multi-task learning and prediction model that separates and extracts coupled information.

[0070] Specifically, the model's total loss function is:

[0071]

[0072] In the formula: L k loss Let a be the loss function for the k-th task; k These are the weights of the k-th loss function. The weights of the loss function are dynamically updated based on the rate of change.

[0073] a k (t)=a k,0 ×γ k (7);

[0074] In the formula: a k (t) represents the weight of the k-th loss function in the t-th iteration; a k,0 For a k The initial value of γ; k The change rate is denoted as .

[0075] Step 4: Train and select specific parameters for the multi-task learning prediction model by separating and extracting coupled information. The Adam optimization algorithm is used to solve for the model parameters, and cross-validation is used to optimize hyperparameters such as learning rate, number of expert networks, number of feature extraction network layers, and number of network layers.

[0076] Comparative Example: A real photovoltaic dataset from a certain region was used for verification. Three similar photovoltaic power plants were selected through correlation analysis, and the method of this invention was applied to predict the performance of these three power plants. The LSTM algorithm was used as a comparison method. The mean absolute percentage error (MAPE) was used to evaluate the prediction accuracy of the method. The prediction results are as follows: the MAPE of the LSTM method for the three photovoltaic power plants were 8.83%, 9.02%, and 14.21%, respectively, while the MAPE of the method of this invention were 7.25%, 8.41%, and 6.98%, respectively. The method of this invention outperforms the comparison method, demonstrating the effectiveness and advancement of the method of this invention.

[0077] 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 variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-station photovoltaic power prediction coupled information separation extraction multi-task learning method, characterized in that, The method comprises the following steps: Step S1: calculating power similarity, calculating meteorological similarity, and calculating geographical position similarity according to power information, meteorological information, and geographical position information of the photovoltaic power station, comprehensively judging the correlation between photovoltaic power stations through comprehensive similarity, and identifying similar power station groups; Step S2: constructing a coupled information separation and extraction multi-task learning prediction model, which comprises an input layer, a feature extraction network, a prediction network, a gate network, and an output layer; The input layer inputs a prediction input feature set of the similar power station group; The feature extraction network comprises a task expert network for extracting unique information of each power station and a shared network for extracting coupled information between multiple power stations; the task expert network and the shared network are composed of multiple fully connected layer neural networks; The gate network selectively fuses the outputs of the task expert network and the shared network; the output of the gate network of the kth task in the jth layer feature extraction network can be expressed as: D k,j (x) = w D k,j (D k,j-1 (x))S k,j (x) (5) In the formula, k represents task number; x is input data; w D k,j For the weight function of task k in the jth layer feature extraction network, the Softmax function is calculated to obtain: w D k,j = Softmax(W k,j x), wherein W k,j is a parameter matrix; The prediction network takes the final output of the gate network as its input feature, which includes both unique information of each photovoltaic power station and coupled information between different power stations; the prediction network is composed of a long short-term memory neural network; The output layer is a power prediction value of each power station; Step 3: constructing a dynamic update loss function of the coupled information separation and extraction multi-task learning prediction model; Step 4: training and selecting specific parameters of the coupled information separation and extraction multi-task learning prediction model.

2. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: The power similarity is calculated by using, for example, a Pearson correlation coefficient to calculate the power similarity between different photovoltaic power stations: where P i and Q i are the power data of two power plants, respectively, and are their mean values.

3. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: The meteorological similarity is calculated by using a multidimensional vector similarity measure to calculate the similarity of meteorological data: where M i , N i are the meteorological data vectors of the two power plants, respectively.

4. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: When calculating the geographical position similarity, the Haversine formula is used to calculate the geographical distance between photovoltaic power stations: Wherein, φ and λ are the latitude and longitude, respectively, and r is the radius of the earth.

5. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: The comprehensive similarity is calculated as follows: S = w1 · S power + w2 · S weather + w3 · S location (4); where S power , S location and S weather are the similarity of power, weather and geographical location respectively, and w1, w2, w3 are their weights.

6. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: In step S1, the similar power station groups are identified, specifically by using a DBSCAN clustering algorithm to group the power stations.

7. The multi-task learning method for coupled information separation and extraction of multi-station photovoltaic power prediction according to claim 1, characterized in that: The prediction input feature set comprises photovoltaic power historical data, a prediction time, irradiance, temperature, humidity, cloud cover, and rainfall.

8. The multi-task learning method for coupled information separation and extraction of multi-site photovoltaic power prediction according to claim 1, characterized in that: The dynamic update loss function of the coupled information separation and extraction multi-task learning prediction model is as follows: In the formula, L k loss Loss function of the kth task; a k Weight of the kth loss function; the weight of the loss function is dynamically updated according to the change rate: a k (t) = a k,0 × γ k (7); In the formula: a k (t) represents the weight of the k-th loss function in the t-th iteration; a k,0 For a k Initial value of γ; k The change rate is denoted as .

9. The multi-task learning method for coupled information separation and extraction of multi-site photovoltaic power prediction according to claim 1, characterized in that: The model parameters are solved by using an Adam optimization algorithm, and the learning rate, the number of expert networks, the number of feature extraction network layers, the number of network layers, and other hyperparameters are optimized and selected by using a cross-validation method.

Images