SCSSA-based photovoltaic power generation prediction method

Abstract

The invention relates to a photovoltaic power generation prediction method based on SCSSA, and the method comprises the steps: collecting the data of a photovoltaic power station, and carrying out the preprocessing of the data, and the data comprise historical power generation data, meteorological data, geographic information, and equipment operation data; performing multi-scale decomposition on the preprocessed data, and optimizing multi-scale feature construction by using an SCSSA algorithm; performing dynamic importance evaluation on the multi-scale features by adopting an improved SHAP method to obtain new features related to a time decay factor; and constructing and using an SCSSA algorithm to optimize an adaptive CNN-BiLSTM model, and using the optimized CNN-BiLSTM model to predict the photovoltaic power generation amount. Compared with the prior art, the method has the advantages of more comprehensively capturing the complex dynamic characteristics of the photovoltaic power generation system, improving the prediction precision and efficiency and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation prediction, and in particular to a photovoltaic power generation prediction method based on SCSSA. Background Art

[0002] As the global demand for clean energy continues to grow, photovoltaic power generation, as an important form of renewable energy, has become increasingly prominent in the energy field. However, the output power of photovoltaic power generation is significantly intermittent and uncertain, which poses severe challenges to the stable operation, planning and scheduling of power systems.

[0003] In recent years, with the development of machine learning technology, a variety of algorithms have been applied to photovoltaic power generation prediction. These methods can be roughly divided into two categories: statistical methods and artificial intelligence methods.

[0004] Statistical methods include time series analysis, regression analysis, etc. These methods rely on the statistical characteristics of historical data, but they often have difficulty capturing complex time series patterns.

[0005] Artificial intelligence-based photovoltaic power generation prediction methods have made significant progress, but still face some challenges:

[0006] 1. Multi-scale time series analysis: Photovoltaic power generation data usually contains changes at multiple time scales, such as hourly, daily and even seasonal changes, which requires the prediction model to capture information at different time scales simultaneously.

[0007] 2. Non-stationarity: Due to changes in weather conditions, photovoltaic power generation data shows obvious non-stationarity, which puts higher requirements on the generalization ability of the model.

[0008] 3. Long-term and short-term dependencies: Both long-term trends and short-term fluctuations in photovoltaic power generation data need to be accurately modeled, which poses a challenge to the model's sequence processing capabilities.

[0009] 4. Feature selection and fusion: How to effectively select the most influential features from a large number of candidate features and integrate them into the model is also the key to improving prediction accuracy.

[0010] 5. Setting empirical parameter values: In the actual use of the model, it is often necessary to adjust multiple hyperparameters. These adjustments are usually based on intuition or experience. Different empirical setting values ​​will affect the final result, and there is often a process of continuous trial and error.

[0011] How to achieve accurate and stable photovoltaic power generation prediction has become a technical problem that needs to be solved. Summary of the invention

[0012] The purpose of the present invention is to provide a photovoltaic power generation prediction method based on SCSSA in order to overcome the defects of the above-mentioned prior art.

[0013] The purpose of the present invention can be achieved by the following technical solutions:

[0014] According to one aspect of the present invention, a photovoltaic power generation prediction method based on SCSSA is provided, the method comprising the following steps:

[0015] Step S1, collecting and preprocessing data of the photovoltaic power station, wherein the data includes historical power generation data, meteorological data, geographic information and equipment operation data;

[0016] Step S2, performing multi-scale decomposition on the pre-processed data and optimizing the multi-scale feature construction using the SCSSA algorithm;

[0017] Step S3, using the improved SHAP method to perform dynamic importance evaluation on the multi-scale features of step S2 to obtain new features related to the time decay factor;

[0018] Step S4, based on the results of steps S1 to S3, construct and use the SCSSA algorithm to optimize the adaptive CNN-BiLSTM model, and use the optimized CNN-BiLSTM model to predict photovoltaic power generation.

[0019] Preferably, the multi-scale decomposition of the pre-processed data is specifically: multi-scale decomposition of the pre-processed data is performed using discrete wavelet transform and empirical mode decomposition.

[0020] Preferably, the optimizing the multi-scale feature construction using the SCSSA algorithm comprises:

[0021] Optimizing the parameters of discrete wavelet transform, including wavelet basis, optimal number of multi-scale decomposition layers, and thresholds for denoising and feature extraction;

[0022] The parameters of the empirical mode decomposition are optimized, wherein the parameters include: the number of iterations of the empirical mode decomposition and the processing method of the time series power outage.

[0023] Preferably, a time decay factor is introduced to improve the SHAP, comprising:

[0024] The formula for calculating the time-weighted SHAP value is:

[0025]

[0026] Where m is the number of samples, φ i,j The SHAP value of feature i in the sample, w(t j ) is the time attenuation factor of feature j at time t,

[0027] The formula for the time decay factor is:

[0028]

[0029] Among them, λ is the decay rate parameter, t0 is the current time, and t is the timestamp of the data point.

[0030] Preferably, the SCSSA algorithm is used to optimize the dynamic importance assessment, including the optimization of the number of feature clusters, the sampling rate and the time decay factor.

[0031] Preferably, the use of the SCSSA algorithm to optimize the CNN-BiLSTM model includes: for each sparrow position, constructing a corresponding CNN-BiLSTM model, iteratively optimizing and selecting the sparrow position with the smallest objective function value as the optimal parameter, and using the optimal parameter to construct the final CNN-BiLSTM model;

[0032] The CNN layer and BiLSTM layer of the CNN-BiLSTM model are used to extract local weather pattern features and process long-term and short-term temporal dependencies, respectively.

[0033] Preferably, the prediction in step S4 is a rolling prediction, that is, the power generation for the next m days is predicted each time, the prediction is updated every n days, and new data of the photovoltaic power station is collected.

[0034] Preferably, step S4 includes using the top-n models obtained in the SCSSA optimization process to perform integrated prediction, taking the average of the prediction values ​​of these models to obtain the final prediction result, calculating the confidence interval of the prediction result, and normalizing the prediction results within the confidence interval.

[0035] Preferably, the method further comprises multi-objective evaluation and feedback after step S4, taking the comprehensive score as the main optimization target of the SCSSA algorithm, and adjusting the search strategy of the SCSSA according to the multi-objective evaluation results;

[0036] The comprehensive scoring function used in the multi-objective evaluation is:

[0037] Score = w1×(1-RMSE n )+w2×(1-MAE n )+w3×R 2 +w4×PICP-w5×C n

[0038] Among them: w1, w2, w3, w4, w5 are weight coefficients; RMSE n is the normalized root mean square error; MAE n is the normalized mean absolute error; R2 is the determination coefficient to measure the goodness of fit of the model; PICP is the prediction interval coverage probability; C n is the normalized model complexity.

[0039] Preferably, the method further includes online learning and model updating after step S4, dynamically updating the training data set through sliding window technology to retrain the CNN-BiLSTM model, and regularly triggering the SCSSA optimization process to achieve continuous optimization of the CNN-BiLSTM model.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1) The present invention constructs multi-scale features, evaluates the dynamic importance of multi-scale features using the improved SHAP method, builds an adaptive CNN-BiLSTM model, and uses SCSSA optimization to output photovoltaic power generation prediction using the CNN-BiLSTM model after global optimization. This can more comprehensively capture the complex dynamic characteristics of the photovoltaic power generation system and significantly improve the prediction accuracy and stability.

[0042] 2) The multi-scale feature construction and improved SHAP method of the present invention for dynamic feature importance evaluation enrich the features of the original data and provide the interpretability of the model decision, which helps to understand the impact of different factors on the prediction results; the adaptive model CNN-BiLSTM structure enables flexible adaptation to different environmental conditions and data pattern changes, has strong adaptability, and enhances the robustness of the prediction.

[0043] 3) The SCSSA of the present invention can automatically search and determine the optimal hyperparameter combination, reduce the influence of subjective factors and the process of manual trial and error, and has a high degree of automation; through an efficient search strategy, the most satisfactory solution can be quickly found, the number of iterations can be reduced, and a large amount of computing resources can be saved, thereby enhancing prediction efficiency.

[0044] 4) The multi-objective evaluation and feedback of the present invention adjusts the search strategy of SCSSA according to the multi-objective evaluation results, enhances the robustness of the system, and makes the prediction results more reliable.

[0045] 5) The online learning and model updating mechanism of the present invention enables the system to continuously adapt to new data and environmental changes, maintaining the long-term effectiveness of the prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the process of the photovoltaic power generation prediction method in the present invention;

[0047] Figure 2 This is the comparison of the prediction results of the test set before SCSSA optimization in the present invention;

[0048] Figure 3 Comparison of test set prediction results after SCSSA optimization in the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0050] The purpose of the present invention is to provide a photovoltaic power generation prediction method based on SCSSA to solve the problems of empirical setting of hyperparameters, influence of subjective factors, insufficient prediction accuracy, poor adaptability, etc. in the prior art. The method is more accurate and comprehensive in photovoltaic power generation prediction to meet the growing needs of the power system and improve the utilization efficiency of photovoltaic power generation and the stability of the power grid.

[0051] The present embodiment relates to a photovoltaic power generation prediction method based on SCSSA, which mainly uses the sparrow search algorithm SCSSA based on the fusion of sine and cosine and Cauchy variation for global optimization, wherein SCSSA initializes the population with the help of the refraction reverse learning mechanism on the basis of the sparrow search algorithm (Sparrow Search Algorithm, SSA), increases species diversity, then introduces the sine and cosine strategy and the global and local optimization ability of the nonlinear decreasing search factor and weight factor coordination algorithm in the update of the discoverer position, and finally introduces the Cauchy variation in the follower position to disturb and update the optimal solution, improve the search efficiency and accuracy of the SSA algorithm, and thus find the optimal solution. Then, by comprehensively using data collection and preprocessing, multi-scale feature construction, dynamic feature importance evaluation, and adaptive CNN-BiLSTM model construction. Finally, the sparrow search algorithm SCSSA fusion of sine and cosine and Cauchy variation is used to globally optimize the dynamic feature importance evaluation and the adaptive CNN-BiLSTM model, and realizes the high-precision prediction of photovoltaic power generation. Combined with integrated learning and uncertainty analysis, the robustness and reliability of the model are further enhanced. This method effectively captures the complex characteristics of photovoltaic power generation at different time scales and provides a scientific basis for the optimal management of photovoltaic power generation systems.

[0052] like Figure 1 , the method comprises the following steps:

[0053] Step S1. Data collection and preprocessing: collect historical power generation data, meteorological data, geographic information and equipment operation data from the photovoltaic power station, and perform outlier processing, missing value filling and normalization processing;

[0054] Step S2. Multi-scale feature construction: Use wavelet transform and empirical mode decomposition (EMD) method to perform multi-scale decomposition on the pre-processed data to capture the features of different time scales;

[0055] Step S3. Dynamic feature importance evaluation: The improved SHAP (SHapley Additive ex Planations) method is used. The improvement is: a time decay factor is introduced to perform dynamic feature importance evaluation to adapt to the time-varying characteristics of the photovoltaic power generation system;

[0056] Step S4. Adaptive CNN-BiLSTM model construction: Design an adaptive prediction model consisting of CNN layer and BiLSTM layer, which is suitable for processing sequence data and extracting temporal and spatial sequence features.

[0057] Step S5. SCSSA optimization control: using the sparrow search algorithm (SCSSA) fused with sine-cosine and Cauchy mutation to globally optimize steps S2 to S4, including optimizing feature extraction parameters, feature importance evaluation parameters and prediction model structure;

[0058] Step S6. Model training and prediction: Use the optimized model parameters for training and perform prediction on the test set;

[0059] Step S7. Multi-objective evaluation and feedback: Calculate multi-dimensional evaluation indicators, including RMSE, MAE, prediction interval coverage, etc., and feed the evaluation results back to step S5 for further optimization;

[0060] Step S8. Online learning and model updating: Dynamically update the training data set through the sliding window technology, and regularly trigger the SCSSA optimization process to achieve continuous optimization of the model;

[0061] Step S9. Uncertainty analysis: Use an ensemble learning method to generate multiple SCSSA-optimized CNN-BiLSTM models, calculate the confidence intervals and probability distributions of the prediction results, and analyze the uncertainties from different sources.

[0062] The purpose of step S1 is to obtain high-quality input data to lay the foundation for subsequent feature extraction and model training, including:

[0063] Step S1.1 Data acquisition

[0064] The following data are collected from the PV plant:

[0065] Historical power generation data: Power generation (kW) recorded at 5-minute intervals.

[0066] Meteorological data: including light intensity (W / m 2), irradiance, temperature (℃), humidity (%), wind speed (m / s), etc. The collection frequency is consistent with the power generation data.

[0067] Geographic information: latitude and longitude coordinates and altitude of the power station.

[0068] Equipment operation data: including the type, installation angle, and service life of PV panels.

[0069] Step S1.2 Data preprocessing

[0070] Outlier processing: The three-time standard deviation method was used to identify outliers. For the identified outliers, if the data of the adjacent time points were normal, they were replaced by linear interpolation; otherwise, they were marked as missing values.

[0071] Missing value processing: For short-term (≤3 hours) missing values, linear interpolation is used to fill them; for long-term missing values, historical data with similar dates and weather conditions are used to fill them.

[0072] Data normalization: Min-Max normalization method is used to map all eigenvalues ​​to [0,1].

[0073] Step S2 aims to capture the changing characteristics of the photovoltaic power generation system at different time scales and improve the prediction model's ability to identify complex patterns, including:

[0074] Step S2.1 Wavelet transform

[0075] The input sequence is decomposed into multiple scales using discrete wavelet transform (DWT). By default, Daubechies wavelet (db4) is selected as the mother wavelet, and three layers of decomposition are performed to obtain three detail coefficient sequences and one approximate coefficient sequence.

[0076] The specific process is as follows:

[0077] 1. Perform high-pass and low-pass filtering on the input sequence x[n].

[0078] 2. Downsample the filtered sequence by a factor of 2.

[0079] 3. Repeat steps 1 and 2 until the specified number of decomposition levels is reached.

[0080] Mathematical expression:

[0081]

[0082] in, is the approximate coefficient, W ψ (j,k) is the detail coefficient, j is the scale parameter, k is the translation parameter, and ψ are the scaling function and wavelet function respectively.

[0083] S2.2 Empirical Mode Decomposition (EMD)

[0084] The EMD method is used to adaptively decompose nonlinear and non-stationary time series. The specific steps are as follows:

[0085] 1. Identify all local extreme points of the signal.

[0086] 2. Generate upper and lower envelopes through spline interpolation.

[0087] 3. Calculate the mean of the envelope: m(t) = (e up (t)+e low (t)) / 2.

[0088] 4. Subtract the mean envelope from the original signal: h(t) = x(t) - m(t).

[0089] 5. Check if h(t) satisfies the conditions of the internal model function (IMF). If not, repeat steps 1-4.

[0090] 6. Subtract the IMF from the original signal and repeat the above steps for the remaining signal until the remaining signal becomes a monotonic function.

[0091] Finally, the original signal is decomposed into several IMFs and a residual term:

[0092]

[0093] Among them, c i (t) is IMF, r n (t) is the residual term.

[0094] The multi-scale features decomposed in steps S2.1 and S2.2 are used as new features.

[0095] S2.3 SCSSA Optimization of Multi-Scale Feature Construction

[0096] In order to find the best combination of DWT and EMD parameters, the extracted features can maximize the performance indicators of classification or regression tasks. The SCSSA algorithm is used to optimize the following parameters:

[0097] DWT parameters include:

[0098] a1. Wavelet basis: Select different wavelet types such as db1, db2, sym4, and coif1 to try and explore whether other wavelet bases are better.

[0099] a2. Number of decomposition layers: Determine the optimal number of multi-scale decomposition layers.

[0100] a3.Threshold: The threshold parameter used for denoising and feature extraction. Threshold processing of detail coefficients can reduce noise.

[0101] EMD parameters include:

[0102] b1. Number of decompositions: The number of iterations of EMD decomposition, which affects the decomposition accuracy.

[0103] b2. Endpoint processing method: How to handle the endpoints of the time series (such as mirror extension, linear extrapolation, etc.) can affect the decomposition results.

[0104] In step S3, an improved SHAP (SHapley Additive exPlanations) method is introduced to dynamically evaluate the importance of features to adapt to the time-varying characteristics of the photovoltaic power generation system. For each feature output in step S2, a time-related weight-adjusted SHAP value sequence is obtained as a new feature.

[0105] The improved SHAP method is as follows:

[0106] c1. Calculate the standard SHAP value:

[0107]

[0108] Among them, N is the set of all features, S is the subset that does not contain feature i, and f x (S) is the model prediction value using only feature subset S.

[0109] c2. Introduce time decay factor:

[0110]

[0111] Among them, λ is the decay rate parameter, t0 is the current time, and t is the timestamp of the data point.

[0112] c3. Calculate the time-weighted SHAP value:

[0113]

[0114] Where m is the number of samples, φ i,j SHAP value of feature i in the ith sample.

[0115] c4.SCSSA optimizes SHAP calculation:

[0116] The SCSSA algorithm is used to optimize the calculation process of the SHAP value, mainly optimizing the following parameters:

[0117] A. Number of feature clusters: Cluster related features to reduce computational complexity.

[0118] B. Sampling rate: controls the sample ratio used for SHAP value calculation.

[0119] C. Time decay factor λ: adjusts the influence of time weight.

[0120] In step S4, an adaptive CNN-BiLSTM model structure is designed, which can effectively capture the spatiotemporal characteristics of photovoltaic power generation data.

[0121] Step S4.1 Model structure design, including

[0122] Input layer: Shape: (batch size, number of features = 10), Description: Input the features processed by steps S1, S2 and S3, including multiple time and space features such as longitude and latitude, light intensity, irradiance, temperature, humidity, etc.

[0123] CNN layer group: used to extract local weather pattern features, including:

[0124] A. The first convolutional layer, number of filters: 64, kernel size: 3, activation function: ReLU.

[0125] B. First maximum pooling layer, pooling size: 2.

[0126] C. Second convolutional layer, number of filters: 128, convolution kernel size: 3, activation function: ReLU.

[0127] D. Second maximum pooling layer, pooling size: 2

[0128] 3. BiLSTM layer group: used to handle long-term and short-term time dependencies

[0129] A. First bidirectional LSTM layer, number of LSTM units: 128, return sequence: yes.

[0130] B. Second bidirectional LSTM layer, number of LSTM units: 64, return sequence: no.

[0131] 4. Fully connected layer group:

[0132] A. The first fully connected layer, number of neurons: 64, activation function: ReLU.

[0133] B. Dropout layer, dropout rate: 0.3.

[0134] C. Output layer, number of neurons: 1, activation function: linear.

[0135] Step S4.2 Model compilation settings

[0136] Loss function: Mean Squared Error (MSE)

[0137] Optimizer: Adam optimizer

[0138] Learning rate: 0.001

[0139] Momentum term β1: 0.9

[0140] Adaptive learning rate term β2: 0.999

[0141] Evaluation indicators: root mean square error (RMSE), mean absolute error (MAE).

[0142] In step S5, it is explained in detail how the SCSSA algorithm optimizes the above CNN-BiLSTM model to improve the prediction accuracy of photovoltaic power generation.

[0143] S5.1 Initialization settings

[0144] Sparrow population size: 30

[0145] Maximum number of iterations: 100

[0146] Sparrow position encoding: [number of CNN layers, number of first layer filters, number of second layer filters, number of BiLSTM layers, number of first layer LSTM units, number of second layer LSTM units, Dropout rate, learning rate, batch size].

[0147] Parameter range settings: Number of CNN layers: 1-3, number of filters: [32, 64, 128, 256], number of BiLSTM layers: 1-2, number of LSTM units: [32, 64, 128, 256], Dropout rate: 0.1-0.5, learning rate: 0.0001-0.01, batch size: [16, 32, 64, 128].

[0148] S5.2 Optimization process, including:

[0149] A. Sparrow position initialization:

[0150] Randomly generate 30 sparrow positions that meet the parameter range

[0151] Example position: [2,64,128,2,128,64,0.3,0.001,64]

[0152] B. Iterative optimization (repeat 100 times or reach early stopping condition):

[0153] a. For each sparrow position, build a corresponding CNN-BiLSTM model, use the historical data of the last month to train the model (10 rounds), verify it on the data of the last 7 days, and calculate the objective function:

[0154] f=0.4RMSE+0.3MAE-0.2R 2 +0.1×model complexity

[0155] b. Update the sparrow position:

[0156] Discoverers (top 20%):

[0157] x 新 =x 旧 +Random number×|x 旧 -x 最佳 |

[0158] Followers (middle 70%):

[0159] x 新 =x 最佳 +Random number×|x 旧 -x 最佳 |

[0160] Vigilant (bottom 10%):

[0161] x 新 =x 旧 +0.1×(X 平均 -x 旧 )

[0162] Where: x 新 is the updated sparrow position (i.e., the new model parameter combination); x 旧 is the current sparrow position (current model parameter combination); x 最佳 is the global optimal position in the current population; the random number refers to a random number uniformly distributed in the interval [-1,1].

[0163] c. Parameter boundary processing:

[0164] Ensure that the updated parameters are within the preset range and truncate or rebound the out-of-bounds parameters.

[0165] d. Dynamic adjustment (every 10 iterations):

[0166] Learning step size: α = α × 0.95

[0167] Global search probability: p = p × 0.98 (initial value 0.8)

[0168] e. Early stop check:

[0169] If the optimal solution does not improve after 10 consecutive iterations, the process is terminated early.

[0170] D. Optimal model selection and application:

[0171] a. Select the sparrow position with the smallest objective function value as the optimal solution;

[0172] b. Use the optimal parameters to build the final CNN-BiLSTM model;

[0173] c. Perform final training on the model using 60 days of complete historical data;

[0174] d. Apply the optimized model to the 9-day power generation forecast.

[0175] The comparison of the prediction results of the test set before and after SCSSA optimization is as follows: Figure 2 and Figure 3 shown.

[0176] In step S6, how to use the SCSSA-optimized CNN-BiLSTM model for training and prediction is described in detail.

[0177] Step S6.1 Model training process

[0178] A. Dataset division:

[0179] Training set: 70%, take the first 42 days of data

[0180] Validation set: 15%, using the last 9 days of data

[0181] Test set: 15%, take the last 9 days of data

[0182] B. Batch training:

[0183] Batch size: set according to SCSSA optimization results

[0184] Epochs: Set to 100, but use early stopping mechanism. Early stopping condition: If the validation set loss does not improve within 10 epochs, stop training.

[0185] C. Learning rate scheduling:

[0186] Initial learning rate: set according to SCSSA optimization results

[0187] Learning rate decay: halve the learning rate every 20 epochs

[0188] Step S6.2 Prediction process

[0189] A. Rolling Forecast:

[0190] Each time the power generation is predicted for the next 9 days, the prediction is updated every 7 days to incorporate new actual data into the input. B. Ensemble learning and uncertainty analysis:

[0191] Use the top-5 models obtained during the SCSSA optimization process for ensemble prediction.

[0192] Compute confidence intervals for the predictions:

[0193] CI=[μ-1.96σ,μ+1.96σ]

[0194] Among them, μ is the average prediction value of the integrated model, and σ is the standard deviation of the prediction value

[0195] Generate the probability distribution of the prediction results, assuming that it follows the normal distribution N(μ,σ2)

[0196] C. Post-processing of prediction results:

[0197] The normalized prediction results are inverted back to the actual power generation, and the prediction values ​​are checked for rationality to ensure they are non-negative. In step S7, it is detailed how to evaluate the model performance and feed the results back to the SCSSA optimization process.

[0198] Step S7.1 Multi-dimensional evaluation indicators

[0199] A. Forecast accuracy indicators:

[0200] Root mean square error (RMSE)

[0201] Mean Absolute Error (MAE)

[0202] Mean Absolute Percentage Error (MAPE)

[0203] Coefficient of determination (R 2 )

[0204] B. Model complexity index:

[0205] Number of parameters: The total number of parameters in the model

[0206] Inference time: Average inference time (milliseconds / sample) on standard hardware

[0207] Step S7.2. Comprehensive scoring mechanism

[0208] Design a comprehensive scoring function that integrates the above indicators into a single score:

[0209] Score = w1×(1-RMSE n )+w2×(1-MAE n )+w3×R 2 +w4×PICP-w5×C n

[0210] in:

[0211] w1, w2, w3, w4, and w5 are weight coefficients that can be adjusted according to specific application requirements. The indicator with _norm added indicates that normalization has been performed;

[0212] RMSE n is the normalized root mean square error;

[0213] MAE n is the normalized mean absolute error;

[0214] R 2 The coefficient of determination measures the goodness of fit of the model;

[0215] PICP is the prediction interval coverage probability, which indicates the reliability of the prediction interval;

[0216] C n is the normalized model complexity, reflecting the simplicity and computational efficiency of the model.

[0217] Step S7.3 Feedback mechanism

[0218] A. Performance report generation:

[0219] Generate detailed performance reports including all evaluation metrics and composite scores, visualize key indicators such as time series plots of predicted power generation vs. actual power generation.

[0220] B.SCSSA optimization feedback:

[0221] The comprehensive score is taken as the main optimization target of the SCSSA algorithm. According to the evaluation results, the search strategy of SCSSA is adjusted, such as increasing the search density within a specific parameter range.

[0222] C. Model update trigger:

[0223] Set a performance threshold to trigger model re-optimization when the comprehensive score is below the threshold, and perform a complete SCSSA optimization regularly (such as monthly) to adapt to potential system changes.

[0224] In step S8, online learning and model updating include:

[0225] Step S8.1 Sliding Window Technique

[0226] A. Window size W: set to 30 days.

[0227] B. Sliding step size S: set to 7 days.

[0228] C. Update method: Each time an update is performed, the earliest S days of data in the window are removed, the latest S days of data are added, and the model is retrained using the updated data set.

[0229] Step S8.2 Incremental Learning

[0230] A. Weight initialization: Keep the current weights of the model as initial values.

[0231] B. Fine-tuning training: Use new data to perform fine-tuning training for a small number of epochs.

[0232] C. Learning rate: Set to 1 / 10 of the original learning rate to prevent overfitting to new data.

[0233] Step S8.3 Regular SCSSA optimization

[0234] A. Optimization cycle T: Optimization is performed every 30 days.

[0235] B.SCSSA algorithm: Re-optimize the model structure and hyperparameters in each optimization cycle

[0236] C. Optimize the objective function:

[0237] f=α×P h +(1-α)×P r

[0238] Among them: α is the balance coefficient, which is used to adjust the weight of historical performance and recent performance;

[0239] P h is a historical performance indicator;

[0240] P r is the performance indicator for the most recent data.

[0241] In step S9, the uncertainty analysis includes:

[0242] Step S9.1 Ensemble learning method

[0243] In order to solve the uncertainty problem in model prediction, an ensemble learning method is used. Specifically, multiple SCSSA-optimized CNN-BiLSTM models are constructed through the Bagging method. Each model is trained using a different Bootstrap sampled data subset. This method helps to improve the robustness and generalization ability of the model.

[0244] In the prediction stage, the outputs of all models are integrated, and the final prediction result is obtained by taking the average of the prediction values ​​of these models. This can smooth out the errors that may be generated by individual models, thereby obtaining more stable and reliable prediction results.

[0245] Step S9.2 Prediction interval calculation

[0246] To quantify the uncertainty of the forecast, we calculate the forecast mean and standard deviation and construct a confidence interval for the forecast based on them. For example, a 95% confidence interval can be used to characterize the forecast range, indicating that in 95% of cases, the actual value will fall within the forecast interval. This interval is constructed based on the output of multiple models, so that it can reflect the overall uncertainty of the forecast.

[0247] Step S9.3 Analysis of sources of uncertainty

[0248] Data uncertainty: mainly comes from the quality and variability of input data. By analyzing the volatility and noise level of the data, the impact of the uncertainty of the data itself on the prediction results can be evaluated.

[0249] Model uncertainty: Since different model structures may produce different results when dealing with the same problem, it is necessary to evaluate model uncertainty by comparing the performance of different models. This analysis can reveal the impact of model structure on prediction stability.

[0250] Parameter uncertainty: During the SCSSA optimization process, the choice and convergence of parameters may affect the final performance of the model. By monitoring and analyzing how the parameters change during the optimization process, the impact of parameter uncertainty on the model results can be evaluated.

[0251] In summary, the present invention proposes a photovoltaic power generation prediction method based on SCSSA, which realizes high-precision prediction of photovoltaic power generation by comprehensively applying data collection and preprocessing, multi-scale feature construction, dynamic feature importance evaluation, adaptive CNN-BiLSTM model construction, and SCSSA optimization control. Integrated learning and uncertainty analysis further enhance the robustness and reliability of the model. This method effectively captures the complex characteristics of photovoltaic power generation at different time scales, provides a scientific basis for the optimal management of photovoltaic power generation systems, and has broad application prospects.

[0252] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A photovoltaic power generation prediction method based on SCSSA, characterized in that: The method comprises the following steps: Step S1, collecting and preprocessing data of the photovoltaic power station, wherein the data includes historical power generation data, meteorological data, geographic information and equipment operation data; Step S2, performing multi-scale decomposition on the pre-processed data and optimizing the multi-scale feature construction using the SCSSA algorithm; Step S3, using the improved SHAP method to perform dynamic importance evaluation on the multi-scale features of step S2 to obtain new features related to the time decay factor; Step S4, based on the results of steps S1 to S3, construct and use the SCSSA algorithm to optimize the adaptive CNN-BiLSTM model, and use the optimized CNN-BiLSTM model to predict photovoltaic power generation.

2. A photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The multi-scale decomposition of the pre-processed data specifically includes: performing multi-scale decomposition of the pre-processed data using discrete wavelet transform and empirical mode decomposition.

3. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The use of the SCSSA algorithm to optimize the multi-scale feature construction includes: Optimizing the parameters of discrete wavelet transform, including wavelet basis, optimal number of multi-scale decomposition layers, and thresholds for denoising and feature extraction; The parameters of the empirical mode decomposition are optimized, wherein the parameters include: the number of iterations of the empirical mode decomposition and the processing method of the time series power outage.

4. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The time decay factor is introduced to improve the SHAP, including: The formula for calculating the time-weighted SHAP value is: Where m is the number of samples, φ i,j The SHAP value of feature i in the sample, w(t j ) is the time attenuation factor of feature j at time t, The formula for the time decay factor is: Among them, λ is the decay rate parameter, t0 is the current time, and t is the timestamp of the data point.

5. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The SCSSA algorithm is used to optimize the dynamic importance assessment, including the optimization of the number of feature clusters, the sampling rate, and the time decay factor.

6. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The use of the SCSSA algorithm to optimize the CNN-BiLSTM model includes: for each sparrow position, constructing a corresponding CNN-BiLSTM model, iteratively optimizing and selecting the sparrow position with the smallest objective function value as the optimal parameter, and using the optimal parameter to construct the final CNN-BiLSTM model; The CNN layer and BiLSTM layer of the CNN-BiLSTM model are used to extract local weather pattern features and process long-term and short-term temporal dependencies, respectively.

7. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The prediction in step S4 is a rolling prediction, that is, the power generation for the next m days is predicted each time, the prediction is updated every n days, and new data of the photovoltaic power station is collected.

8. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The step S4 includes using the top-n models obtained in the SCSSA optimization process to perform integrated prediction, calculating the average of the prediction values ​​of these models to obtain the final prediction result, calculating the confidence interval of the prediction result, and normalizing the prediction result within the confidence interval.

9. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The method further includes multi-objective evaluation and feedback after step S4, taking the comprehensive score as the main optimization target of the SCSSA algorithm, and adjusting the search strategy of the SCSSA according to the multi-objective evaluation results; The comprehensive scoring function used in the multi-objective evaluation is: Score=w1×(1-RMSE n )+w2×(1-MAE n )+w3×R 2 +w4×PICP-w5×C n Among them: w1, w2, w3, w4, w5 are weight coefficients; RMSE n is the normalized root mean square error; MAE n is the normalized mean absolute error; R 2 is the determination coefficient to measure the goodness of fit of the model; PICP is the prediction interval coverage probability; C n is the normalized model complexity.

10. The photovoltaic power generation prediction method based on SCSSA according to claim 1, characterized in that: The method also includes online learning and model updating after step S4, dynamically updating the training data set through the sliding window technology to retrain the CNN-BiLSTM model, and regularly triggering the SCSSA optimization process to achieve continuous optimization of the CNN-BiLSTM model.

Images