Photovoltaic power generation power prediction method based on deep learning and multi-source data fusion

Abstract

The invention discloses a photovoltaic power generation power prediction method based on deep learning and multi-source data fusion, and the method comprises the steps: carrying out the preprocessing of the real-time power generation data, environment parameter data and weather forecast data of a photovoltaic power station, reducing the data noise through Gaussian filtering and bilateral filtering, and then carrying out the prediction of the real-time power generation data, environment parameter data and weather forecast data; according to the method, feature extraction is carried out on data based on a model of an LSTM and an Attention mechanism, the LSTM is used for capturing a long and short term dependency relationship in a time sequence, the Attention mechanism focuses on key environment features through a weighting mode, the feature extraction precision and the model interpretation are improved, finally, multi-source data are fused into a final power prediction result by using a full-connection layer, and the power prediction accuracy is improved. And further smoothing processing is carried out through median filtering, so that the stability and the anti-interference performance of a prediction result are improved. According to the method, the real-time performance and accuracy of photovoltaic power prediction can be improved under the complex environment condition, and the adaptability and robustness of the system are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the fields of smart grid and deep learning, and specifically discloses a photovoltaic power generation power prediction method based on deep learning and multi-source data fusion. Background Art

[0002] With the rapid development of new energy and smart grid technologies, photovoltaic power generation has become an important part of the energy system due to its clean and renewable characteristics. However, since photovoltaic power generation is affected by external factors such as weather and environmental conditions, the power output has large volatility and uncertainty, which poses severe challenges to grid dispatching and power supply. Therefore, how to accurately predict photovoltaic power generation has become a research focus in the current power system field. Existing photovoltaic power generation prediction methods mainly rely on historical power generation data and simple environmental parameter information, and are modeled through traditional machine learning algorithms (such as support vector machines, regression analysis, etc.) or shallow neural networks. However, these methods have significant shortcomings, especially under complex and changeable weather conditions.

[0003] Traditional prediction methods have poor adaptability to environmental changes and are difficult to effectively capture the medium- and long-term time dependence of photovoltaic power generation. The use of a single data source is incapable of dealing with complex nonlinear relationships, resulting in low prediction accuracy. Especially under conditions of sudden weather changes or long-term climate change, the model is difficult to effectively reflect future power generation trends. In addition, traditional methods have limitations in multi-source data fusion, and cannot fully integrate multi-dimensional information such as power station data, environmental data, and weather forecast data. It is impossible to fully reflect the various factors that affect power generation, which limits the reliability and application scope of the model.

[0004] With the rapid development of deep learning technology, time series data processing methods based on long short-term memory (LSTM) networks and attention mechanisms have gradually shown their advantages in photovoltaic power forecasting. LSTM networks are particularly suitable for processing photovoltaic power generation time series data due to their outstanding performance in modeling long-term and short-term dependencies; the attention mechanism can dynamically assign weights according to the importance of the data, strengthen the focus on key environmental features, and further improve the accuracy and interpretability of the forecast. In addition, multi-source data fusion technology makes it possible to effectively combine power station data, environmental parameters, and weather forecast information, thereby building a more forward-looking and stable photovoltaic power forecasting model.

[0005] At present, there is an urgent need for a photovoltaic power prediction method that combines multi-source data fusion and deep learning technology to overcome the shortcomings of traditional methods, improve the accuracy and real-time performance of predictions, and meet the needs of high-precision photovoltaic power prediction in the fields of smart grid scheduling, energy management, and new energy optimization configuration. The focus is on developing a new prediction method that combines LSTM with the Attention mechanism and fuses multi-source data to effectively capture the impact of environmental changes on power output and enhance the stability and robustness of the model. Summary of the invention

[0006] The present invention relates to the field of new energy and smart grid technology, and discloses a photovoltaic power prediction method based on deep learning and multi-source data fusion. The technical problem solved by the method is how to efficiently fuse and predict multi-source data under complex environmental conditions to improve the accuracy and real-time performance of photovoltaic power prediction and meet the needs of grid dispatching and new energy management.

[0007] S1. Data preprocessing: Use Gaussian filtering and bilateral filtering to reduce noise on the collected raw photovoltaic power station data and environmental parameter data to improve data quality and stability. In Gaussian filtering, the standard deviation σ is set to 1.5, and the filter window radius k is set to 3.5 to smooth the data and reduce the impact of random fluctuations. The data after Gaussian filtering is obtained. In bilateral filtering, the spatial domain standard deviation σ s Set to 3.0, grayscale standard deviation σ r Set to 0.7, filter window radius k to 8.0, and get the bilateral filtered data bilateral_data. Bilateral filtering retains key feature information while smoothing, ensuring that the denoised data still retains important edge and detail features, thereby improving the accuracy of subsequent predictions.

[0008] S2. Multi-source data fusion: Through multi-source data fusion technology, the real-time power data of photovoltaic power stations, environmental parameter data (such as irradiance, temperature, humidity, wind speed, etc.) and weather forecast data are integrated into a multi-dimensional feature group and normalized. The specific process is: divide the multi-source data into adjacent intervals according to the time series, extract the maximum value of the feature and its corresponding time point in each interval, and form a feature group {P_i, T_i}, where i = 1, 2, ..., n, i represents the interval number, P_i represents the maximum value of the feature in the i-th interval, and T_i represents the time point corresponding to the maximum value of the feature in the i-th interval. This process effectively enhances the stability of the data, reduces the impact of noise, and ensures that the extracted key features can fully reflect the real dynamics of photovoltaic power changes with the environment, thereby providing high-quality data input for subsequent feature extraction and prediction.

[0009] S3. Feature extraction: The fused multi-source data is processed by dimension upscaling and input into the deep learning model combining LSTM and Attention to extract local and global features in the time series. First, the long-term and short-term dependencies in the data are modeled through a multi-layer LSTM structure. Each LSTM layer captures the continuous change pattern and complex dependency characteristics in the time series through recursive calculation. The specific process includes inputting the dimension-upgraded multi-source data into the LSTM network to generate implicit states reflecting short-term and long-term trends, and further extracting local features of power fluctuations over time. The Attention mechanism is applied to the time series features output by the LSTM. The Attention mechanism automatically identifies the most critical environmental features and moments for prediction by calculating the importance weights of each time step. The process includes a multi-head self-attention mechanism and weight allocation, which focuses the model's "attention point" on specific time points and features, such as critical moments when irradiance changes significantly. The Attention mechanism improves the model's sensitivity to the global features of the data by weighting different features, while enhancing the ability to capture long-term dependencies and ensuring the effective extraction of key information.

[0010] S4, Model training and optimization: Use the data collected in real time from power stations and environmental parameters to conduct preliminary training of LSTM and Attention mechanism models, and gradually adjust the model weights through iterative optimization to enable it to dynamically adapt to complex meteorological and environmental conditions. The specific process includes using historical data for initial training and real-time data updates for dynamic fine-tuning of the model to ensure the balance between prediction accuracy and response speed of the model.

[0011] S5. Dual prediction: After the model training is completed, the system can receive the latest power generation data and environmental parameters (such as real-time irradiance, temperature, humidity, wind speed, etc.) of the photovoltaic power station in real time, and immediately use the trained model to make short-term power predictions. Through the LSTM model, the system captures short-term fluctuations and change trends in the input data, so that the model can accurately predict short-term changes in power generation within the time range of minutes to hours, providing rapid response support for power grid dispatch. The system also combines weather forecast data to make medium- and long-term predictions of future power. By inputting future weather forecast data (such as irradiance, temperature, humidity, wind speed, etc. for a period of time in the future), the model uses a combination of LSTM and Attention to capture long-term power change trends. LSTM is responsible for establishing the dependency relationship of the power time series, while the Attention mechanism enables the model to focus on the time points and features in the weather forecast data that have the greatest impact on power fluctuations, thereby more accurately predicting future power trends. This dual prediction method is not only suitable for the short-term dispatch needs of the power grid, but also can provide data support for long-term power generation plans.

[0012] S6, feature smoothing: Perform median filtering on the predicted power curve to further remove possible abnormal noise points and improve the smoothness and accuracy of the predicted data. Median filtering can retain important power generation trend information when processing real-time prediction curves, ensuring that the power grid dispatching system can make reasonable dispatching decisions based on stable data.

[0013] Through the above steps, the method of the present invention can effectively integrate multi-source data in a complex environment, capture the dynamic changes of photovoltaic power generation, improve the accuracy and real-time performance of the prediction, and is suitable for fields that require high-precision power prediction, such as smart grid scheduling, smart power station management, and new energy optimization configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic diagram of a photovoltaic power prediction method based on deep learning and multi-source data fusion provided in an embodiment of the present invention.

[0015] Figure 2 It is a structural diagram of the multi-head attention mechanism provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] like Figure 1 and Figure 2 As shown, this embodiment provides a photovoltaic power prediction method based on deep learning and multi-source data fusion, comprising the following steps:

[0017] S1. Data preprocessing: Use Gaussian filtering and bilateral filtering to reduce noise on the collected raw photovoltaic power station data and environmental parameter data to improve data quality and stability. In Gaussian filtering, the standard deviation σ is set to 1.5, and the filter window radius k is set to 3.5 to smooth the data and reduce the impact of random fluctuations. The data after Gaussian filtering is obtained. In bilateral filtering, the spatial domain standard deviation σ s Set to 3.0, grayscale domain standard deviation σ r Set to 0.7, filter window radius k to 8.0, and get the bilateral filtered data bilateral_data. Bilateral filtering retains key feature information while smoothing, ensuring that the denoised data still retains important edge and detail features, thereby improving the accuracy of subsequent predictions.

[0018] S2. Multi-source data fusion: Through multi-source data fusion technology, the real-time power data, environmental parameter data and weather forecast data of the photovoltaic power station are integrated into a multi-dimensional feature group and normalized. The specific process is: divide the multi-source data into adjacent intervals according to the time series, extract the maximum value of the feature and its corresponding time point in each interval, and form a feature group {P_i, T_i}, where i = 1, 2, ..., n, i represents the interval number. This process effectively enhances the stability of the data, reduces the impact of noise, and ensures that the extracted key features can fully reflect the real dynamics of photovoltaic power changes with the environment, thereby providing high-quality data input for subsequent feature extraction and prediction.

[0019] S3. Feature extraction: The fused multi-source data is processed by dimension upscaling and input into the deep learning model combining LSTM and Attention to extract local and global features in the time series. First, the long-term and short-term dependencies in the data are modeled through a multi-layer LSTM structure. Each LSTM layer captures the continuous change pattern and complex dependency characteristics in the time series through recursive calculation. The specific process includes inputting the dimension-upgraded multi-source data into the LSTM network to generate implicit states reflecting short-term and long-term trends, and further extracting local features of power fluctuations over time. The Attention mechanism is applied to the time series features output by the LSTM. The Attention mechanism automatically identifies the most critical environmental features and moments for prediction by calculating the importance weights of each time step. The process includes a multi-head self-attention mechanism and weight allocation, which focuses the model's "attention point" on specific time points and features, such as critical moments when irradiance changes significantly. The Attention mechanism improves the model's sensitivity to the global features of the data by weighting different features, while enhancing the ability to capture long-term dependencies and ensuring the effective extraction of key information.

[0020] S4, Model training and optimization: Use the data collected in real time from power stations and environmental parameters to conduct preliminary training of LSTM and Attention mechanism models, and gradually adjust the model weights through iterative optimization to enable it to dynamically adapt to complex meteorological and environmental conditions. The specific process includes using historical data for initial training and real-time data updates for dynamic fine-tuning of the model to ensure the balance between prediction accuracy and response speed of the model.

[0021] S5. Dual prediction: After the model training is completed, the system can receive the latest power generation data and environmental parameters (such as real-time irradiance, temperature, humidity, wind speed, etc.) of the photovoltaic power station in real time, and immediately use the trained model to make short-term power predictions. Through the LSTM model, the system captures short-term fluctuations and change trends in the input data, so that the model can accurately predict short-term changes in power generation within the time range of minutes to hours, providing rapid response support for power grid dispatch. The system also combines weather forecast data to make medium- and long-term predictions of future power. By inputting future weather forecast data (such as irradiance, temperature, humidity, wind speed, etc. for a period of time in the future), the model uses a combination of LSTM and Attention to capture long-term power change trends. LSTM is responsible for establishing the dependency relationship of the power time series, while the Attention mechanism enables the model to focus on the time points and features in the weather forecast data that have the greatest impact on power fluctuations, thereby more accurately predicting future power trends. This dual prediction method is not only suitable for the short-term dispatch needs of the power grid, but also can provide data support for long-term power generation plans.

[0022] S6, feature smoothing: Perform median filtering on the predicted power curve to further remove possible abnormal noise points and improve the smoothness and accuracy of the predicted data. Median filtering can retain important power generation trend information when processing real-time prediction curves, ensuring that the power grid dispatching system can make reasonable dispatching decisions based on stable data.

[0023] Next, the six steps are explained one by one.

[0024] Step S1: Data preprocessing

[0025] The step S1 specifically includes two parts:

[0026] S11, Gaussian filtering: First, read the original data file of the photovoltaic power station and environmental parameters and convert it into a three-dimensional data array of size (150,360,288), recorded as X. Then use Gaussian filtering to process X and obtain the smoothed data X gaussian The Gaussian filter formula is as follows:

[0027]

[0028] Where x, y are the coordinates within the filter window; σ is the standard deviation, which is set to 1.5.

[0029] Use Gaussian filtering to perform convolution on three-dimensional data X:

[0030]

[0031] Where r is the radius of the filter window, which is set to 3; X(i,j,k) is the value of the original data at the coordinate (i,j,k); X gaussian (i, j, k) is the value of the data after Gaussian filtering at the coordinate (i, j, k).

[0032] S12, bilateral filtering: Next, use bilateral filtering to filter X gaussian Processing is performed to retain important edge features. In bilateral filtering, the standard deviation σ in the spatial domain s Set to 3.0, the standard deviation of the grayscale domain σ r Set to 0.7, and the radius r of the filter window is set to 8.0. The formula for bilateral filtering is as follows:

[0033]

[0034] Among them, G s (a, b, c) is a Gaussian function in the spatial domain, defined as:

[0035]

[0036] Gaussian weight G in spatial domain r (ΔX) is:

[0037]

[0038] Where ΔX = X gaussian (i+a,j+b,k+c)-X gaussian (i, j, k) is the grayscale difference between the data points; and W(i, j, k) is the normalization factor to ensure that the filtered data maintains numerical consistency, defined as:

[0039]

[0040] Bilateral filtering combines information from the spatial domain and the smoothing domain. While reducing noise, it can effectively retain edge features and details in the data, further improving the robustness of the data.

[0041] Through the noise reduction processing of Gaussian filtering and bilateral filtering, smooth photovoltaic power station and environmental parameter data with high signal-to-noise ratio are finally obtained, which enhances the stability of the data and lays the foundation for the prediction and analysis of subsequent models.

[0042] S2. Multi-source data fusion: Through multi-source data fusion technology, we integrate data from different sources, and after interval division, feature extraction and normalization, finally form a multi-dimensional feature group. This process effectively enhances the stability and robustness of the data and ensures the correlation between different data sources, thereby providing higher quality data support for subsequent predictive modeling. The specific process is as follows:

[0043] S21. Data division and feature extraction: divide the real-time power data, environmental parameter data and weather forecast data of the photovoltaic power station into several adjacent time intervals according to the time series. Assume that the time series data is {D1, D2, …, D N}, where N is the total number of data. The data is divided into n adjacent time periods by interval, and the interval number is i (i = 1, 2, ..., n). The length of each interval is Δt, that is, each interval contains data of Δt consecutive time points. In each interval, the maximum value of the feature and its corresponding time point are extracted from the power data, environmental parameter data and weather forecast data respectively. Let the feature of each data source be P i (Photovoltaic power station power), E i (environmental parameters) and W i (weather forecast data), then the feature group of each interval {P i ,T i} is given by the following formula:

[0044] P i =max (7)

[0045] T i = arg max (8)

[0046] Among them, P i represents the power data in the i-th interval, T i is the maximum power P in this interval i The corresponding time point. Similarly, the environmental data E i and weather forecast data W i The maximum value can also be extracted similarly:

[0047] S22, data normalization and fusion: The obtained feature group {P i ,E i ,W i} is normalized, and Min-Max normalization is used to scale each feature to the range of [0,1]. The specific formula is as follows:

[0048]

[0049]

[0050]

[0051] Among them, P min and P max Respectively represent the minimum and maximum values ​​of power data, E min and E max Respectively represent the minimum and maximum values ​​of environmental parameter data, Wmin and W max Respectively represent the minimum and maximum values ​​of weather forecast data.

[0052] S23, data fusion and output: All normalized feature data are integrated into a multi-dimensional feature group to form the final input data:

[0053]

[0054] S3: The fused multi-source data is processed for dimensionality increase and input into the deep learning model combining LSTM and Attention to extract local and global features in the time series. This process mainly includes the following steps:

[0055] S31. Dimensionality Upgrading: Based on the fusion of multi-source data, the fused data is firstly processed to upgrade its dimension to meet the input requirements of LSTM and Attention models. Assume that the fused feature data is X i , whose dimension is d. We upgrade the data to a higher dimension d′ to capture more time series features and input them into the deep learning model.

[0056] The dimension increase process can be achieved through the fully connected layer. The specific method is as follows:

[0057]

[0058] Where f represents the dimension-raising function, W expand is the weight matrix of the dimension-raising operation, This is the data after dimensionality upgrade.

[0059] S32, LSTM network modeling: Input the multi-source data after dimensionality increase into the LSTM network to model the long-term and short-term dependencies in the time series data. LSTM recursively calculates through multiple LSTM layers to capture the continuous change patterns and complex time dependencies in the sequence data. The process is as follows:

[0060] Input time series data It is processed through a series of LSTM layers. Each LSTM layer takes the hidden state of the previous layer as input and updates the current hidden state by calculation to capture short-term and long-term trends. The output of the LSTM network is a sequence H i , which contains the hidden state representation for each time step. The calculation formula of the LSTM network is as follows:

[0061]

[0062] Among them, H i-1 is the hidden state of the previous time step, is the input of the current time step, H i is the output of the current time step.

[0063] S33, Multi-Head Self-Attention Mechanism: First, a Multi-Head Self-Attention mechanism is used to parallelize the weights of different time steps. Each attention head focuses on different aspects of the time series features, allowing the model to learn richer representations. The specific process is as follows:

[0064] S331. Input division: The output sequence of LSTM H = [H1, H2, ..., H n ] as input, with a vector dimension of d for each time step.

[0065] S332. Linear transformation: for the input sequence H i Use three sets of weight matrices W Q , W K , W V Perform linear transformation to generate query vector Q, key vector K and value vector V:

[0066] Q i =W Q ·H i ,K i =W K ·H i ,V i =W V ·H i (15)

[0067] S333. Similarity calculation: Perform dot product on Q and K of each time step to get the attention score score (H i ,H j ):

[0068]

[0069] Among them, d k is the dimension of the key vector, Used for scaling to avoid dot product values ​​being too large.

[0070] S334. Weight normalization: Use the Softmax function to normalize the similarity score to obtain the attention weight α for each time step i :

[0071]

[0072] S335. Weighted aggregation output: For each time step, multiply its hidden state by the attention weight to obtain the weighted feature, and sum the weighted features of all time steps to obtain the global feature representation

[0073]

[0074] Finally, the global features output by the Attention mechanism It is a comprehensive representation that captures important information in time series data. It effectively strengthens the features of key time steps and ignores minor or irrelevant moments, thereby improving the predictive power of the model.

[0075] S4. Model training and optimization: Entering the model training and optimization stage, the goal is to train, fine-tune and optimize the deep learning model that combines LSTM and Attention by combining historical data and real-time data.

[0076] S41. Model training:

[0077] S411, forward propagation: input the training set features into the model:

[0078]

[0079] Among them, X i = {P i ,E i ,W i}.

[0080] S412. Loss calculation: Calculate the loss based on the predicted value and the actual value:

[0081]

[0082] S413, Back Propagation: Update model parameters using gradient descent:

[0083]

[0084] Among them, θ represents the parameters of the model.

[0085] S414, validation set evaluation: After each epoch, use the validation set to evaluate the model performance, calculate the validation loss Val_Loss, and record it to check for overfitting.

[0086] S42. Real-time data update and dynamic fine-tuning: The model is dynamically adjusted through the introduction of real-time data to adapt to the real-time changing environment.

[0087] S421, Real-time data collection and processing: Processing real-time data into feature input

[0088]

[0089] in, Specifically, it is used to collect power data, environmental data and environmental data of photovoltaic power stations.

[0090] S422, incremental training: Use the sliding window mechanism to fine-tune the most recent T data and update the weights synchronously:

[0091] X t-T:t ={X t ,X t-1 ,…,X t-T+1} (twenty three)

[0092]

[0093] Among them, θ represents the parameters of the model; η low is the learning rate, set to 0.0001 to avoid over-adjustment.

[0094] S5. Dual prediction: After the model is trained and optimized, the system can realize short-term and medium-term dual power prediction based on real-time data and future weather forecast data. By combining LSTM and Attention mechanisms, short-term fluctuations and long-term trends are modeled respectively, providing all-round data support for power grid dispatch and power generation planning.

[0095] S51. Short-term power prediction: using LSTM network Processing is performed to capture short-term fluctuations and changing trends in time series, and accurate predictions are made on power changes ranging from minutes to hours.

[0096]

[0097] in, Represents the short-term power forecast value; After preprocessing and feature extraction by S1 and S2, a real-time feature sequence is formed.

[0098] S52. Medium- and long-term power forecast: Combined with future weather forecast data, it captures the long-term trend of power output changes and provides support for power generation plans in the next few hours or even days.

[0099] S521. Data input: input future weather forecast data After the data is processed by S1 and S2, a feature sequence is generated

[0100] S522, Model prediction: Use the deep learning model that combines LSTM and Attention to predict.

[0101] S5221, LSTM network is responsible for capturing the dependencies of power time series in weather forecast data:

[0102]

[0103] S5222, Attention mechanism adjusts the model's focus on key features according to the importance of each time point in the weather forecast data:

[0104]

[0105]

[0106] in, Indicates the medium- and long-term power forecast value.

[0107] S6, feature smoothing: use median filtering to further remove isolated noise points and improve data quality. i For each data point in , use the median value in the sliding window to replace the original value, remove noise points and smooth the power curve. The median filter operation is as follows:

[0108] P i ′=median(P i-k ,P i-k+1 ,…,P i+k ) (29)

[0109] Among them, P i ′ is the power data after median filtering, P i is the original prediction value, and k is the radius of the sliding window.

Claims

1. A photovoltaic power prediction method based on deep learning and multi-source data fusion, characterized in that: The following steps are involved: S1. Data preprocessing: Use Gaussian filtering and bilateral filtering to reduce noise on the collected original photovoltaic power station data and environmental parameter data to improve data quality and stability; S2. Multi-source data fusion: Through multi-source data fusion technology, the real-time power data, environmental parameter data and weather forecast data of the photovoltaic power station are integrated into a multi-dimensional feature group and normalized; S3: Feature extraction: The fused multi-source data is processed for dimensionality increase and input into the deep learning model combining LSTM and Attention to extract local and global features in the time series; S4, model training and optimization: Use the data collected in real time from power stations and environmental parameters to conduct preliminary training of LSTM and Attention mechanism models, and gradually adjust the model weights through iterative optimization to enable it to dynamically adapt to complex meteorological and environmental conditions; S5: Dual prediction: After the model training is completed, the system can receive the latest power generation data and environmental parameters of the photovoltaic power station in real time, and immediately use the trained model to perform short-term power prediction; S6, feature smoothing: Perform median filtering on the predicted power curve to further remove possible abnormal noise points and improve the smoothness and accuracy of the predicted data.

2. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 1 is characterized in that: The environmental parameter data include irradiance, temperature, humidity, and wind speed.

3. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 1 is characterized in that: In step S1, the standard deviation σ of the Gaussian filter is set to 1.5, and the filter window radius k is set to 3.

5. In the bilateral filter, the spatial domain standard deviation σ s Set to 3.0, grayscale standard deviation σ r is set to 0.7, and the filter window radius k is set to 8.

0.

4. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 1 is characterized in that: In step S2, the multi-source data is divided into adjacent intervals according to the time series, and the feature maximum value and its corresponding time point are extracted in each interval to form a feature group {P_i, T_i}, where i = 1, 2, ..., n, i represents the interval number, P_i represents the feature maximum value of the i-th interval, and T_i represents the time point corresponding to the feature maximum value of the i-th interval.

5. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 1 is characterized in that: Step S3 specifically includes: S31, Dimensionality Upgrading: Based on the fusion of multi-source data, the fused data is firstly processed to increase its dimension to meet the input requirements of LSTM and Attention models; S32, LSTM network modeling: The multi-source data after dimensionality increase is input into the LSTM network to model the long-term and short-term dependencies in the time series data. LSTM recursively calculates through multiple LSTM layers to capture the continuous change patterns and complex time dependencies in the sequence data; S33, Multi-Head Self-Attention Mechanism: First, a multi-head self-attention mechanism is used to parallelize the weights of different time steps. Each attention head focuses on different aspects of the time series features, enabling the model to learn richer representations.

6. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 5 is characterized in that: The specific process in step S33 includes: S331: Input division: The output sequence of LSTM H = [H1, H2, ..., H n ] as input, the vector dimension of each time step is d; S332: Linear transformation: for the input sequence H i Use three sets of weight matrices W Q , W K , W V Perform linear transformation to generate query vector Q, key vector K and value vector V: Q i =W Q ·H i ,K i =W K ·H i ,V i =W V ·H i S333: Similarity calculation: perform dot product on Q and K of each pair of time steps to obtain the attention score score (H i ,H j ): Among them, d k is the dimension of the key vector, Used for scaling to avoid dot product values ​​being too large; S334: Weight normalization: Use the Softmax function to normalize the similarity score to obtain the attention weight α for each time step i : S335: Weighted aggregation output: For each time step, multiply its hidden state by the attention weight to obtain the weighted feature, and sum the weighted features of all time steps to obtain the global feature representation Finally, the global features output by the Attention mechanism It is a comprehensive representation that can capture important information in time series data.

7. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 6 is characterized in that: Step S4 specifically includes: S41. Model training: S411, forward propagation: input the training set features into the model: Among them, X i = {P i ,E i ,W i }; S412. Loss calculation: Calculate the loss based on the predicted value and the actual value: S413, Back Propagation: Update model parameters using gradient descent: Among them, θ represents the parameters of the model; S414, Validation set evaluation: After each epoch, use the validation set to evaluate the model performance, calculate the validation loss Val_Loss, and record it to check for overfitting; S42, Real-time data update and dynamic fine-tuning: Dynamically adjust the model through the introduction of real-time data to adapt it to the real-time changing environment; S421, Real-time data collection and processing: Processing real-time data into feature input in, Specifically, it is used to collect power data, environmental data and environmental data of photovoltaic power stations; S422, incremental training: Use the sliding window mechanism to fine-tune the most recent T data and update the weights synchronously: X t-T:t ={X t ,X t-1 ,…,X t-T+1 } Among them, θ represents the parameters of the model; η low is the learning rate, set to 0.0001 to avoid over-adjustment.

8. The photovoltaic power generation prediction method based on deep learning and multi-source data fusion according to claim 7 is characterized in that: Step S5 specifically includes: S51. Short-term power prediction: using LSTM network Processing is performed to capture short-term fluctuations and changing trends in time series, and accurate predictions of power changes in the range from minutes to hours are made; in, Represents the short-term power forecast value; After preprocessing and feature extraction by S1 and S2, a real-time feature sequence is formed; S52, medium- and long-term power forecast: Combined with future weather forecast data, it captures the long-term trend of power output changes and provides support for power generation plans in the next few hours or even days; S521. Data input: input future weather forecast data After the data is processed by S1 and S2, a feature sequence is generated S522, Model Prediction: Use the deep learning model combining LSTM and Attention to predict; S5221, LSTM network is responsible for capturing the dependencies of power time series in weather forecast data: S5222, Attention mechanism adjusts the model's focus on key features according to the importance of each time point in the weather forecast data: in, Indicates the medium- and long-term power forecast value.

Images