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

Abstract

The invention belongs to the technical field of power generation power prediction, and particularly relates to a photovoltaic power prediction method based on multi-source data fusion and deep learning, and the method comprises the steps: firstly obtaining the historical power, satellite irradiance and numerical weather forecast data of a photovoltaic power station; performing preprocessing such as space-time alignment, missing value processing and abnormal value elimination, and constructing periodic time features; training a deep learning model taking a long short-term memory network as a core by using the processed data so as to capture a complex nonlinear time sequence relationship between the weather and the power; inputting the satellite and forecast data in a to-be-predicted time period into the model, and outputting a future multi-step power predicted value; and finally, carrying out online deviation correction and physical amplitude limiting post-processing to obtain a final prediction result. By effectively fusing multi-source data and deep learning, the precision and practicability of short-term photovoltaic power prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power generation power prediction, and particularly relates to a photovoltaic power prediction method and system based on multi-source data fusion and deep learning. BACKGROUND

[0002] With the transformation of global energy structure towards clean and low-carbon, the installed capacity and grid-connected proportion of photovoltaic power generation continue to grow rapidly. Photovoltaic power has the characteristics of intermittency, volatility and randomness, and its large-scale grid connection poses a severe challenge to the safe and stable operation and real-time economic dispatch of the power system. Under this background, high-precision short-term photovoltaic power prediction (usually referring to prediction in the future several hours to several days) has become a key supporting technology to ensure power grid consumption, optimize power station operation and participate in power market transactions.

[0003] At present, the short-term photovoltaic power prediction methods can be mainly divided into physical model method, statistical model method and data-driven method based on machine learning / deep learning.

[0004] The physical model method is based on the physical characteristics of photovoltaic components, the geometric position of the sun and meteorological parameters (mainly irradiance and temperature), and the theoretical power generation power is calculated by establishing a photoelectric conversion model. This method has clear physical meaning, but has significant limitations in practical application: first, its prediction accuracy is highly dependent on the accuracy of input meteorological data, and ground meteorological observation sites are sparse and numerical weather prediction has inherent errors, especially in complex weather with rapid changes in cloud cover, the uncertainty of irradiance input will cause the prediction result to deviate significantly from the actual value; second, the model is usually based on ideal conditions and cannot accurately describe the performance degradation and nonlinear loss caused by factors such as dust accumulation, component aging and local shading, and has poor adaptability.

[0005] The statistical model method (such as autoregressive moving average model, exponential smoothing method, etc.) mainly extrapolates the prediction based on the time series rules of historical power data itself. This method is simple to calculate and can better capture the daily cycle, seasonal cycle and other rules of power. However, its essence is a linear or weakly nonlinear model, which lacks the ability to describe power fluctuations affected by severe weather changes and cannot effectively integrate multi-dimensional meteorological information such as temperature, cloud cover and wind speed, so it often has large prediction deviations when the weather changes suddenly.

[0006] Methods based on traditional machine learning, such as support vector regression, random forest and gradient boosting decision tree, can handle nonlinear relationships and introduce multi-feature inputs, which are better than pure statistical methods. However, these methods usually treat time series as independent and identically distributed sample points for processing, and fail to fully exploit the inherent temporal dependence and long / short-term context information in the data, limiting their modeling capabilities for continuous and dynamically evolving meteorological-power processes. In addition, their feature engineering and model generalization capabilities also face challenges.

[0007] In recent years, deep learning methods, especially recurrent neural networks and their variants such as long short-term memory networks, have shown potential in photovoltaic power prediction due to their powerful temporal modeling capabilities. However, existing deep learning-based methods still suffer from the following common problems: First, at the data level, they often rely on a single data source (such as numerical weather prediction or historical power data alone), failing to fully integrate the advantages of high spatiotemporal resolution satellite real-time observation data and physically consistent numerical weather prediction data. Satellite data can capture real-time changes in local cloud clusters, but it has no forecasting capability for future periods; numerical weather prediction can provide physical field forecasts for the next few days, but it may have biases at short-term initial moments. Effective fusion of the two is key to improving prediction accuracy, but existing methods lack systematic and efficient solutions for spatiotemporal alignment, quality control, and feature fusion of multi-source heterogeneous data. Second, at the model application level, the prediction results often lack systematic bias correction and physical rationality constraints in post-processing, which may lead to predicted values ​​exceeding the actual possible range, affecting their direct usability in scheduling decisions.

[0008] Therefore, there is an urgent need for a short-term photovoltaic power prediction method that can effectively integrate multi-source data from real-time satellite observations and numerical weather predictions, utilize advanced deep learning models to capture complex nonlinear time-series relationships, and possess automated data preprocessing, intelligent feature engineering, and robust post-processing capabilities. This method aims to overcome the shortcomings of existing technologies and achieve more accurate, reliable, and practical prediction results. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a photovoltaic power prediction method and system based on multi-source data fusion and deep learning. By fusing high spatiotemporal resolution satellite irradiance data from Fengyun-4B, ECMWF numerical weather prediction data, and historical power data from photovoltaic power plants, a multi-source fusion input feature set is constructed. This effectively combines the real-time nature of satellite observations, the forecasting capability of NWP data, and the historical operating patterns of power plants. Practical verification shows that this method improves prediction accuracy under complex meteorological conditions such as cloudy skies and sudden weather changes.

[0010] The technical solution adopted by this invention to solve its technical problem is: to provide a photovoltaic power prediction method based on multi-source data fusion and deep learning, comprising the following steps:

[0011] S1. Obtain historical power generation data, meteorological satellite irradiance data, and meteorological data of the target photovoltaic power station;

[0012] S2. Perform spatiotemporal alignment and preprocessing on historical power generation data, meteorological satellite irradiance data, and meteorological data to generate a multi-source fusion input dataset with unified spatiotemporal resolution.

[0013] S3. Construct and train a deep neural network model based on the multi-source fusion input dataset. The model is used to learn the nonlinear mapping relationship between meteorological features and photovoltaic power.

[0014] S4. Input the satellite remote sensing meteorological data and ECMWF data corresponding to the time period to be predicted into the trained deep neural network model, and output the photovoltaic power prediction value within the future set time window;

[0015] S5. Perform post-processing correction on the photovoltaic power prediction value to obtain the final short-term photovoltaic power prediction result.

[0016] Furthermore, in step S1:

[0017] The meteorological satellite irradiance data mentioned are surface solar radiation data provided by the Fengyun-4B satellite.

[0018] The meteorological data includes at least several of the following elements: wind speed and direction at different altitudes, surface air pressure, air temperature and dew point temperature at 2 meters altitude, total precipitation, cloud base height, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, direct radiation, diffuse radiation, downward shortwave radiation and its accumulation, downward longwave radiation and its accumulation, and total downward shortwave radiation.

[0019] Furthermore, the spatiotemporal alignment and preprocessing described in step S2 specifically include:

[0020] S21. Data Alignment: Based on the timestamp of the historical power data of the photovoltaic power station, the satellite irradiance data and meteorological data are aligned to the geographical location and time series of the power station through time matching and spatial interpolation.

[0021] S22. Missing value handling: Short-term consecutive missing values ​​in the data are filled using linear interpolation or nearest neighbor interpolation.

[0022] S23. Outlier Detection and Removal: Based on the regression relationship between historical power and satellite irradiance, a residual model is constructed, and samples with residuals exceeding a set threshold are marked as outliers and removed.

[0023] S24. Data Standardization: Standardize numerical features to eliminate the influence of dimensions.

[0024] Furthermore, after generating the multi-source fusion input dataset, step S2 also includes a feature engineering step:

[0025] S25. Time feature construction: Extract the basic fields of year, month, day and hour from the timestamp, and perform sine and cosine transformations on the hour and month to generate periodic time coding features.

[0026] S26. Meteorological feature screening: Screen out numerical features that are strongly correlated with photovoltaic power prediction from the meteorological data.

[0027] Further, in step S2, the spatial interpolation specifically involves: generating a spatial sampling grid with a preset grid resolution centered on the latitude and longitude coordinates of the photovoltaic power station, and interpolating meteorological data to each point of the grid; the grid resolution and size can be dynamically adjusted according to the spatial resolution of the forecast product.

[0028] Furthermore, in step S3, the deep neural network model is a model based on a long short-term memory network, and its construction and training include:

[0029] S31. Data serialization: The multi-source fusion input dataset is constructed into an input sequence and a corresponding power output sequence according to a set time window;

[0030] S32. Model Structure: Construct a network containing at least one LSTM layer. The LSTM layer captures temporal dependencies. The hidden state at the intermediate time step is selected and concatenated with the input features at the corresponding time step. The concatenation is then input into the fully connected layer to generate the predicted value.

[0031] S33. Model Training: Using the error between predicted power and actual power as the loss function, the optimizer is used for iterative training, and the optimal model is saved by evaluating it through the validation set.

[0032] Further, the model training process described in step S33 is as follows: the root mean square error (RMSE) between the predicted power and the true power is used as the loss function; the Adam optimizer is used for iterative training, with the initial learning rate set to 0.001-0.005 and the number of iterations set to 50-100 rounds; 5-fold cross-validation is used during training, and the model performance is evaluated by the RMSE of the validation set, and the model parameters with the best performance on the validation set are saved.

[0033] Furthermore, in step S4, the deep neural network model is used to output power prediction values ​​for multiple consecutive time steps in the future at once.

[0034] Further, in step S5, the post-processing correction includes:

[0035] Using the real-time power data of the photovoltaic power station, a deviation correction model is constructed. The deviation correction model is a linear regression model, with the model prediction value as the independent variable and the difference between the real-time monitoring value and the prediction value as the dependent variable. The prediction value of the future time step is systematically calibrated according to the deviation correction model. At the same time, if the prediction result shows a power value less than 0 or greater than the rated power of the power station, it is corrected to 0 and the rated power of the power station, respectively.

[0036] This invention also provides a system for photovoltaic power prediction based on multi-source data fusion and deep learning, comprising:

[0037] The multi-source data acquisition module is used to acquire historical power data, meteorological satellite irradiance data, and ECMWF meteorological data of the target photovoltaic power station.

[0038] The data fusion preprocessing module is used to perform spatiotemporal alignment, missing value processing, outlier removal and standardization on the multi-source data, and to perform feature engineering to generate the model input feature set;

[0039] The deep learning model module is used to build and train LSTM-based deep neural network models and use these models to perform multi-step prediction of photovoltaic power.

[0040] The prediction result correction and output module is used to perform post-processing correction on the original prediction values ​​of the model and output the final short-term photovoltaic power prediction results.

[0041] The present invention has the following beneficial effects:

[0042] (1) This invention constructs a multi-source fusion input feature set by integrating high spatiotemporal resolution satellite irradiance data from Fengyun-4B satellite, ECMWF numerical weather prediction data, and historical power data from photovoltaic power plants. This effectively combines the real-time nature of satellite observations, the forecasting capability of NWP data, and the historical operating patterns of power plants. Practical verification shows that this method improves prediction accuracy under complex meteorological conditions such as cloudy skies and sudden weather changes.

[0043] (2) This invention employs a deep neural network model based on long short-term memory networks, which can automatically learn and capture the complex nonlinear temporal dependencies between meteorological features and photovoltaic power. Compared with traditional statistical models and shallow machine learning models, LSTM is better at handling dynamic changes in long-term series and has better adaptability and predictive sensitivity to instantaneous weather fluctuations.

[0044] (3) The model of this invention adopts a "multi-step direct prediction" architecture, which outputs the power prediction values ​​for multiple consecutive time steps in the future at one time. This method avoids the error accumulation problem caused by rolling prediction, ensures the consistency of the prediction sequence in time, and is more in line with the power system dispatching requirements for prediction results for continuous time periods.

[0045] (4) This invention constructs an integrated system that includes data acquisition, fusion preprocessing, model training and inference, result correction and output. The system supports automated pipeline operation, has high concurrency processing capability and good scalability, and can meet the business needs of actual power plants for high-frequency, real-time and stable short-term power prediction. Attached Figure Description

[0046] Figure 1 This is a flowchart of a photovoltaic power prediction method based on multi-source data fusion and deep learning according to the present invention;

[0047] Figure 2 This is a flowchart of the spatiotemporal alignment and preprocessing of a photovoltaic power prediction method based on multi-source data fusion and deep learning according to the present invention.

[0048] Figure 3 This is a flowchart illustrating the training process of a deep neural network model for a photovoltaic power prediction method based on multi-source data fusion and deep learning, as described in this invention.

[0049] Figure 4 This is a structural diagram of a photovoltaic power prediction system based on multi-source data fusion and deep learning according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1: Taking a centralized photovoltaic power station with an installed capacity of 100MW in Northwest my country as an example, a short-term power forecast for the next 72 hours (3 days) is conducted. This power station has been operating stably for more than 3 years and has accumulated complete historical power generation data.

[0052] A photovoltaic power prediction method based on multi-source data fusion and deep learning, such as Figure 1 As shown, it includes the following steps:

[0053] S1. Obtain historical power generation data, meteorological satellite irradiance data, and meteorological data of the target photovoltaic power station;

[0054] Historical power generation data: obtained directly from the target photovoltaic power plant's Supervisory Control and Data Acquisition (SCADA) system or Energy Management System (EMS).

[0055] Content and Requirements: Obtain at least 2-3 years of historical active power data, preferably with a time resolution of 15 minutes or 1 hour. The data should include a timestamp (year-month-day hour:minute:second) and the corresponding total power output of the power plant (unit: kW or MW). The data should be as complete as possible, with coverage preferably exceeding 95%.

[0056] Meteorological satellite irradiance data:

[0057] Source: Data products from the Fengyun-4B geostationary meteorological satellite provided by the National Satellite Meteorological Center. The specific product is "Surface Solar Radiation" data.

[0058] Content and Characteristics: This data provides total surface irradiance covering China and surrounding areas, with a spatial resolution of approximately 4-5 kilometers and a temporal resolution of up to 15 minutes. Its advantage lies in its ability to capture real-time changes in cloud cover, compensating for the scarcity of ground-based observation stations.

[0059] Acquisition method: Download gridded data covering the geographical location (latitude and longitude range) of the target power station via data interface or offline file.

[0060] Numerical weather forecast (NWP) data:

[0061] Source: ERA5 reanalysis data (used for historical model training) and High Resolution Forecasts (HRES) products (used for future forecasts) from the European Centre for Medium-Range Weather Forecasts (ECMWF). Alternatively, data from models such as GRAPES from the China Meteorological Administration (CMA) may be used as an alternative or supplement.

[0062] Acquire surface solar radiation data. The meteorological satellite irradiance data is surface solar radiation data provided by Fengyun-4B satellite;

[0063] The meteorological data includes at least several of the following elements: wind speed and direction at different altitudes, surface air pressure, air temperature and dew point temperature at 2 meters altitude, total precipitation, cloud base height, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, direct radiation, diffuse radiation, downward shortwave radiation and its accumulation, downward longwave radiation and its accumulation, and total downward shortwave radiation.

[0064] Surface / near-surface elements: 2-meter air temperature (t2m), 2-meter dew point temperature (d2m), surface air pressure (sp), 10-meter wind speed (u10, v10), total precipitation (tp).

[0065] Radiation-related elements: net shortwave solar radiation (SSR) at the Earth's surface, direct solar radiation at the Earth's surface (FDIR), diffuse solar radiation at the Earth's surface (FDIF), downward shortwave radiation at the Earth's surface (SSRD) and its accumulation, downward longwave radiation at the Earth's surface (STRD) and its accumulation, total cloud cover (TCC), low cloud cover (LCC), middle cloud cover (MCC), and high cloud cover (HCC).

[0066] Upper-altitude elements: wind speed and direction at altitudes of 100 meters, 150 meters, etc., and data on different pressure layers in the troposphere.

[0067] Spatiotemporal resolution: Historical reanalysis data is typically 1 hour long, with a spatial resolution of approximately 0.25° × 0.25°. Forecast data can have a temporal resolution of 1-3 hours, and the forecast duration should cover the forecast requirements (e.g., 72 hours).

[0068] Data storage recommendations: Store all acquired raw data in a time-series database (such as InfluxDB) or a large distributed file system, and create a metadata index for fast retrieval.

[0069] S2. Spatiotemporal alignment and preprocessing are performed on historical power generation data, meteorological satellite irradiance data, and meteorological data to generate a multi-source fusion input dataset with unified spatiotemporal resolution. The goal of this step is to transform heterogeneous, multi-source, and multi-scale raw data into a clean, aligned, high-quality dataset that can be used for model training.

[0070] like Figure 2 As shown, S21, data alignment: Based on the timestamp of the historical power data of the photovoltaic power station, the satellite irradiance data and meteorological data are aligned to the geographical location and time series of the power station through time matching and spatial interpolation.

[0071] Time alignment: Each timestamp of the historical power data of the photovoltaic power plant (e.g., every 15 minutes) is used as the reference time. Data records in satellite data and NWP data with timestamps closest to this reference time are aligned to the precise reference time using linear interpolation. For forecast data, the forecast time is used directly.

[0072] Spatial Alignment: Determining the Center and Grid: Use the latitude and longitude coordinates of the photovoltaic power station (e.g., (E101.25°, N38.75°)) as the center point. Determine a suitable interpolation grid based on the original resolution of the input NWP data (e.g., 0.25° for ERA5). For example, create a 3×3 grid with a grid point spacing of the original resolution (0.25°), covering an area of ​​approximately 0.5°×0.5° around the power station.

[0073] Interpolation operation: For each type of NWP data (e.g., t2m, u10), extract the original grid data covering that grid range. Then, use bilinear interpolation or inverse distance weighted (IDW) interpolation methods to calculate the meteorological element values ​​at the precise location of the photovoltaic power station. For FY-4B satellite data, due to its higher resolution, the value of the pixel where the power station is located can be directly selected, or the values ​​of neighboring pixels can be averaged.

[0074] Output: After spatiotemporal alignment, each timestamp t corresponds to a unified feature vector X. t Includes: power plant power P tAligned satellite irradiance G sat,t And the aligned values ​​of each meteorological element (t2m) t u10 t v10 t ,...).

[0075] S22. Missing value handling: Short-term consecutive missing values ​​in the data are filled using linear interpolation or nearest neighbor interpolation.

[0076] For short-term data loss caused by transmission and storage, linear interpolation is used to fill in the missing data.

[0077] For non-continuous or slightly longer missing data, time series prediction methods or regression imputation using other highly correlated features can be used. In practice, if a certain time period has severe missing data, it may be considered to directly remove samples from that time period.

[0078] S23. Outlier Detection and Removal: Based on the regression relationship between historical power and satellite irradiance, a residual model is constructed, and samples with residuals exceeding a set threshold are marked as outliers and removed.

[0079] 1. Calculate the theoretical maximum irradiance (sunny day irradiance) at each moment. You can use a physical model (such as the Ineichen model) or select the highest percentile (such as 99.9%) of irradiance in historical data as an approximation.

[0080] 2. Establish a simple linear or nonlinear regression relationship between historical power and measured satellite irradiance to obtain the "expected power".

[0081] 3. Calculate the residual: Residual = Actual power - Expected power.

[0082] 4. Statistically analyze the residual series and mark samples with absolute residual values ​​exceeding 3 times the standard deviation (or identified as outliers based on box plots) as outliers. These outliers may originate from power plant shutdowns, maintenance, snow accumulation, dust obstruction, or data errors, and should be removed.

[0083] S24. Data Standardization: Standardize numerical features to eliminate the influence of dimensions.

[0084] Z-Score normalization was employed. For each numerical feature (e.g., power, temperature, wind speed, and radiation), its mean (μ) and standard deviation (σ) were calculated over the entire training dataset, and then transformed: x norm =(x-μ) / σ.

[0085] Note: The mean μ and standard deviation σ are calculated only from the training set and these parameters are saved for the same standardization transformation on the validation set, test set, and future real-time prediction data to avoid data leakage.

[0086] S25. Time feature construction: Extract the basic fields of year, month, day and hour from the timestamp, and perform sine and cosine transformations on the hour and month to generate periodic time coding features.

[0087] Extract basic features from the timestamp: year, month, day, hour, day of the week, and whether it is a holiday.

[0088] Periodic encoding: To address boundary issues related to periodic features such as months and hours, for example, 23:59 and 00:01 have large numerical differences but similar practical meanings, a sine-cosine transform is applied.

[0089] month_sin = sin(2 * π * month / 12),month_cos = cos(2 * π * month / 12)

[0090] hour_sin = sin(2 * π * hour / 24), hour_cos = cos(2 * π * hour / 24)

[0091] Similarly, the "day of the year" can be coded to reflect seasonal cycles.

[0092] S26. Meteorological Feature Screening: Numerical features strongly correlated with photovoltaic power prediction are screened from the meteorological data. Not all meteorological features are equally effective. Feature importance assessment methods (such as tree-based feature importance ranking or mutual information methods) are used for screening.

[0093] The processed feature set is then input into a Random Forest regression model for initial fitting. The contribution of each feature to reducing model impurity is calculated. Typically, features such as total irradiance (satellite or NWP), direct radiation, diffuse radiation, cloud cover, and temperature are found to be the most important, while certain upper-level wind or longwave radiation features are less important. The top 15-20 most important features can be retained to reduce model complexity and prevent overfitting.

[0094] Finally, the selected meteorological features, constructed time features, and satellite irradiance features are combined to form the final model input feature vector.

[0095] Step S3: Construction and training of deep neural network model; This invention preferably uses Long Short-Term Memory (LSTM) network as the core model to effectively capture the long-term and short-term spatiotemporal dependencies in meteorological and power sequences.

[0096] like Figure 3As shown, S31, data serialization: construct the multi-source fusion input dataset into an input sequence and a corresponding power output sequence according to a set time window;

[0097] The flattened time series data generated in step S2 is reconstructed into the sequence samples required for supervised learning. A look-back window T and a forecast horizon H are set.

[0098] Given T = 24 hours, H = 12 hours, and a time resolution of 1 hour, a training sample is constructed as follows:

[0099] Input sequence X input : [X (t-T+1) ,X (t-T+2) ,...,X t ], which is the feature vector of 24 time steps from time t back 23 hours to time t.

[0100] Output sequence Y label :[P (t+1) ,P (t+2) ,...,P (t+H) [], which refers to the power values ​​over 12 time steps from the next 1 hour to the next 12 hours.

[0101] The sliding window method generates a large number of overlapping sequence samples, which constitute the final training set, validation set and test set (usually divided in a ratio of 7:2:1 or 6:2:2).

[0102] S32. Model Structure: Construct a network containing at least one LSTM layer. The LSTM layer captures temporal dependencies. The hidden state at an intermediate time step is selected and concatenated with the corresponding input features, then input into a fully connected layer to generate the predicted value. A multi-layer LSTM neural network is constructed. A typical architecture is as follows: Figure 2 As shown (illustrative):

[0103] 1. Input layer: Receives a tensor of shape (batch_size, T, n_features), where n_features is the dimension of the feature vector.

[0104] 2. LSTM Layers (1-3 layers can be stacked): The first LSTM layer can be set to units=128 or 256, and return_sequences=True to pass the hidden state of each time step to the next layer or subsequent processing. The LSTM layer automatically learns the complex dynamics within the sequence.

[0105] 3. Attention Mechanism: An attention layer is added after the LSTM layer, enabling the model to focus on the most critical moments in the historical sequence for predicting power at a certain future time, such as the impact of irradiance changes a few hours after sunrise on afternoon power.

[0106] 4. Feature Concatenation and Fully Connected Layers: An effective strategy is to not only use the output of the last time step of the LSTM, but also to concatenate the historical hidden states (or attention-weighted context vectors) corresponding to the prediction time with the NWP forecast features corresponding to the prediction time. This is because for the future time t+k, we already have the NWP forecast values ​​for that time (such as temperature and forecasted radiation), and this information is crucial for predicting the power at t+k.

[0107] The concatenated feature vector is input into one or more fully connected (dense) layers for nonlinear transformation, and finally outputs the predicted values ​​for the next H time steps at once through an output layer (dense layer with H neurons). This achieves multi-step direct forecasting, which, as described in claim 8, helps maintain the temporal consistency of the predicted sequence.

[0108] 5. Activation Functions: LSTM internally uses tanh and sigmoid. Fully connected layers commonly use ReLU or its variants (such as LeakyReLU). Output layers typically do not use activation functions (linear activation) because the power values ​​are continuous real numbers.

[0109] S33. Model Training: The error between predicted and actual power is used as the loss function. Iterative training is performed using an optimizer, and the optimal model is saved through validation set evaluation. Loss Function: The root mean square error (RMSE) is used as the loss function and the main evaluation metric. RMSE penalizes larger errors more severely, meeting the strict control requirements of power systems for extreme prediction errors. During training, minimizing RMSE is equivalent to minimizing the mean square error (MSE).

[0110] Optimizer: The Adam optimizer is used, which combines the advantages of momentum and adaptive learning rate. The initial learning rate is set to 0.001. A learning rate scheduler (ReduceLROnPlateau) is also used, which halves the learning rate when the validation set loss no longer decreases over five consecutive epochs.

[0111] Further, the model training process described in step S33 is as follows: the root mean square error (RMSE) between the predicted power and the true power is used as the loss function; the Adam optimizer is used for iterative training, with the initial learning rate set to 0.001-0.005 and the number of iterations set to 50-100 rounds; 5-fold cross-validation is used during training, and the model performance is evaluated by the RMSE of the validation set, and the model parameters with the best performance on the validation set are saved.

[0112] Training process:

[0113] Initialization: Randomly initialize the model parameters.

[0114] Iterative training: Set the total number of training rounds to 80. Set the batch size to 32 or 64.

[0115] Cross-validation: A 5-fold cross-validation strategy is used to more robustly evaluate model performance and select hyperparameters. The training set is divided into 5 parts, and 4 parts are used for training and 1 part for validation in turn, which is repeated 5 times. The average validation error is taken as the model performance estimate.

[0116] Regularization and overfitting prevention: The following strategies are used during training:

[0117] Early stopping: Monitor the validation set RMSE. If it no longer decreases for 15 consecutive epochs, stop training and roll back to the model parameters with the best performance on the validation set.

[0118] Dropout: Adding a Dropout layer between LSTM layers or after a fully connected layer randomly discards a certain percentage of neurons to prevent the network from becoming overly reliant on specific features.

[0119] Model saving: Saves the model weight file (.h5 or .pth format) that performs best on the independent test set throughout the entire cross-validation process.

[0120] Step S4: Model Prediction

[0121] Furthermore, in step S4, the deep neural network model is used to output power prediction values ​​for multiple consecutive time steps in the future at once.

[0122] Input preparation: For the start time t to be predicted now (Usually at the current moment), two parts of data need to be prepared:

[0123] Historical sequence: Collection from t now -T+1 to t now Actual observational (or high-quality analytical) data, including power plant power, aligned satellite data (up to the most recently available time), and corresponding NWP analysis field data, are used. This data undergoes the exact same preprocessing and standardization as during the training phase.

[0124] Future forecast sequence: obtaining from t now +1 to t now +H is the NWP forecast data. Meanwhile, since there is no real-time satellite data for future times, this feature can be set to zero in the input or replaced with the NWP radiation forecast value, consistent with the way future steps are handled during training.

[0125] Input concatenation: The processed historical sequence and the future forecast sequence are concatenated in chronological order to form a complete sequence with the shape (1, T, n). features The input tensor is NWP. The "historical power" feature of future time steps is unknown and can be filled with the power of the most recent time step or 0, but the model mainly relies on the NWP feature for future prediction.

[0126] Model inference: The prepared input tensor is loaded into the trained LSTM model. The model performs forward propagation, outputting the standardized power prediction value Y for the next H time steps at once. pred,norm .

[0127] Destandardization: Using the power mean and standard deviation saved during training, Y... pred,norm Inverse transformation to actual power value Y pred (Unit: MW).

[0128] Further, in step S5, the post-processing correction includes:

[0129] Using the real-time power data of the photovoltaic power station, a deviation correction model is constructed. The deviation correction model is a linear regression model, with the model prediction value as the independent variable and the difference between the real-time monitoring value and the prediction value as the dependent variable. The prediction value of the future time step is systematically calibrated according to the deviation correction model. At the same time, if the prediction result shows a power value less than 0 or greater than the rated power of the power station, it is corrected to 0 and the rated power of the power station, respectively.

[0130] This step aims to correct systematic biases in the model and ensure the physical plausibility of the predictions.

[0131] Bias correction:

[0132] 1. Construct an online bias correction model. Collect the model's rolling predictions (predictions for past moments) and corresponding actual observations for the most recent N hours (e.g., N=24).

[0133] 2. Establish a univariate linear regression model: Δ = a * P pred +b, where Δ=P actual -P pred The parameters a and b are fitted using the least squares method.

[0134] 3. For the newly obtained future prediction sequence Y pred Applying this correction model: Y pred,corrected =Y pred +(a*Y pred +b). This can correct for forecast offsets caused by changes in weather patterns, slow degradation of power plant performance, or systematic biases in NWPs.

[0135] Physical limit:

[0136] Traversing the corrected prediction sequence Y pred , corrected :

[0137] If any value is less than 0, it is set to 0 (the photovoltaic power station does not generate electricity at night or when there is no light).

[0138] If any value is greater than the power plant's rated installed capacity (100MW in this example), then it is set to 100MW. This is the physical upper limit for power generation.

[0139] Final output: Y after post-processing pred , corrected This provides the final, reliable short-term photovoltaic power forecast, which can output the forecast curve for the next 72 hours at 15-minute or 1-hour intervals and provide it to the grid dispatch center or power plant operator.

[0140] Example 2:

[0141] This invention also provides a system for photovoltaic power prediction based on multi-source data fusion and deep learning, comprising:

[0142] The multi-source data acquisition module is used to acquire historical power data, meteorological satellite irradiance data, and ECMWF meteorological data from the target photovoltaic power station. It is responsible for scheduled communication with the SCADA / EMS system, the National Satellite Meteorological Center data platform, and the ECMWF / CMA data service API. It enables automatic data download, parsing, and raw data storage. It features breakpoint resume and error retry mechanisms.

[0143] The data fusion preprocessing module performs spatiotemporal alignment, missing value handling, outlier removal, and standardization on the multi-source data, and performs feature engineering to generate the model input feature set. It provides a configurable pipeline, allowing users to define data processing steps and parameters through configuration files. Integrated feature engineering functionality automatically constructs temporal features and performs feature filtering.

[0144] The deep learning model module is used to build and train LSTM-based deep neural network models and use these models for multi-step prediction of photovoltaic power. The training platform provides a graphical or scripted environment for model building, training, hyperparameter tuning, and evaluation. It integrates with TensorFlow or PyTorch frameworks. The inference service encapsulates the trained model as a microservice, which receives preprocessed input data and returns power prediction values ​​in real time. It supports high concurrency and low latency prediction requests.

[0145] The prediction result correction and output module is used to perform post-processing correction on the model's original prediction values ​​and output the final short-term photovoltaic power prediction results. It implements an online deviation correction algorithm and updates the correction model parameters periodically (e.g., hourly). It executes post-processing rules such as physical limiting. It manages the prediction results, stores them in a database, and generates visual charts and reports. It provides a data interface to push the prediction results to the upper-level scheduling system.

[0146] Task scheduling and monitoring module: Uses a workflow engine or scheduled task framework to coordinate the automated execution of the entire prediction process. Monitors the running status of each module, data freshness, model performance, and prediction error, and issues alarms when anomalies occur.

[0147] Through the above specific implementation, this invention integrates satellite real-time data with NWP forecasts and combines the powerful sequence modeling capabilities of LSTM, significantly improving prediction accuracy under complex scenarios such as cloudy skies and sudden weather changes. In the embodiments, the standardized RMSE (nRMSE) of the 72-hour prediction on the test set can be reduced by 15%-25% compared to a single data source model. The system's data preprocessing and feature engineering stages effectively address data quality issues, and the post-processing correction stage further smooths out systematic errors, ensuring the stability and reliability of the prediction results.

[0148] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on multi-source data fusion and deep learning, characterized in that, Includes the following steps: S1. Obtain historical power generation data, meteorological satellite irradiance data, and meteorological data of the target photovoltaic power station; S2. Perform spatiotemporal alignment and preprocessing on historical power generation data, meteorological satellite irradiance data, and meteorological data to generate a multi-source fusion input dataset with unified spatiotemporal resolution. S3. Construct and train a deep neural network model based on the multi-source fusion input dataset. The model is used to learn the nonlinear mapping relationship between meteorological features and photovoltaic power. S4. Input the satellite remote sensing meteorological data and ECMWF data corresponding to the time period to be predicted into the trained deep neural network model, and output the photovoltaic power prediction value within the future set time window; S5. Perform post-processing correction on the photovoltaic power prediction value to obtain the final short-term photovoltaic power prediction result.

2. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step S1: The meteorological satellite irradiance data mentioned are surface solar radiation data provided by the Fengyun-4B satellite. The meteorological data includes at least several of the following elements: wind speed and direction at different altitudes, surface air pressure, air temperature and dew point temperature at 2 meters altitude, total precipitation, cloud base height, total cloud cover, low cloud cover, middle cloud cover, high cloud cover, direct radiation, diffuse radiation, downward shortwave radiation and its accumulation, downward longwave radiation and its accumulation, and total downward shortwave radiation.

3. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, The spatiotemporal alignment and preprocessing described in step S2 specifically include: S21. Data Alignment: Based on the timestamp of the historical power data of the photovoltaic power station, the satellite irradiance data and meteorological data are aligned to the geographical location and time series of the power station through time matching and spatial interpolation. S22. Missing value handling: Short-term consecutive missing values ​​in the data are filled using linear interpolation or nearest neighbor interpolation. S23. Outlier Detection and Removal: Based on the regression relationship between historical power and satellite irradiance, a residual model is constructed, and samples with residuals exceeding a set threshold are marked as outliers and removed. S24. Data Standardization: Standardize numerical features to eliminate the influence of dimensions.

4. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, Step S2, after generating the multi-source fusion input dataset, also includes a feature engineering step: S25. Time feature construction: Extract the basic fields of year, month, day and hour from the timestamp, and perform sine and cosine transformations on the hour and month to generate periodic time coding features. S26. Meteorological feature screening: Screen out numerical features that are strongly correlated with photovoltaic power prediction from the meteorological data.

5. A photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1 or 3, characterized in that, In step S2, the spatial interpolation specifically involves generating a spatial sampling grid with a preset grid resolution, centered on the latitude and longitude coordinates of the photovoltaic power station, and interpolating meteorological data to each point of the grid. The grid resolution and size can be dynamically adjusted according to the spatial resolution of the forecast product.

6. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step S3, the deep neural network model is a model based on a long short-term memory network, and its construction and training include: S31. Data serialization: The multi-source fusion input dataset is constructed into an input sequence and a corresponding power output sequence according to a set time window; S32. Model Structure: Construct a network containing at least one LSTM layer. The LSTM layer captures temporal dependencies. The hidden state at the intermediate time step is selected and concatenated with the input features at the corresponding time step. The concatenation is then input into the fully connected layer to generate the predicted value. S33. Model Training: Using the error between predicted power and actual power as the loss function, the optimizer is used for iterative training, and the optimal model is saved by evaluating it through the validation set.

7. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 6, characterized in that, The model training process described in step S33 is as follows: the root mean square error between the predicted power and the actual power is used as the loss function; the Adam optimizer is used for iterative training, with the initial learning rate set to 0.001-0.005 and the number of iterations set to 50-100 rounds; Five-fold cross-validation is used during training. The model performance is evaluated by the RMSE of the validation set, and the model parameters with the best performance on the validation set are saved.

8. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step S4, the deep neural network model is used to output power prediction values ​​for multiple consecutive time steps in the future at once.

9. The photovoltaic power prediction method based on multi-source data fusion and deep learning according to claim 1, characterized in that, In step S5, the post-processing correction includes: Using the real-time power data of the photovoltaic power station, a deviation correction model is constructed. The deviation correction model is a linear regression model, with the model prediction value as the independent variable and the difference between the real-time monitoring value and the prediction value as the dependent variable. The prediction value of the future time step is systematically calibrated according to the deviation correction model. At the same time, if the prediction result shows a power value less than 0 or greater than the rated power of the power station, it is corrected to 0 and the rated power of the power station, respectively.

10. A system for implementing the photovoltaic power prediction method based on multi-source data fusion and deep learning as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to acquire historical power data, meteorological satellite irradiance data, and ECMWF meteorological data of the target photovoltaic power station. The data fusion preprocessing module is used to perform spatiotemporal alignment, missing value processing, outlier removal and standardization on the multi-source data, and to perform feature engineering to generate the model input feature set; The deep learning model module is used to build and train LSTM-based deep neural network models and use these models to perform multi-step prediction of photovoltaic power. The prediction result correction and output module is used to perform post-processing correction on the original prediction values ​​of the model and output the final short-term photovoltaic power prediction results.

Images