Zone area net load prediction method considering photovoltaic

Abstract

The invention relates to a zone area net load prediction method considering photovoltaic. The prediction method comprises the following steps: model training; the method specifically comprises the following steps: collecting net load power and meteorological variable data of a fixed area and a fixed interval duration; performing modal decomposition on the net load power by applying geometric modal decomposition to the SGMD; the net load power is calculated, mutual information between the decomposition mode of the net load power and meteorological variables is calculated, and an SGCs strong correlation predictive factor screening model based on the mutual information is constructed to determine the optimal predictive input of different SGCs; a net load SGCs combination prediction model is constructed based on an extrusion excitation network, and model training is completed for different SGCs and prediction factors; and predicting the corresponding regional net load by using the model trained in the previous step. According to the method, the extrusion excitation network is introduced to enhance the attention of the model on important features, the generalization ability of the model is enhanced, and finally the prediction accuracy and stability are improved.

Description

Technical Field

[0001] The present invention relates to digital data processing technology, and particularly to a method for predicting the net load of a distribution area considering photovoltaic power. Background Art

[0002] With the construction of a new power system, the penetration rate of distributed renewable energy represented by photovoltaic power generation in the low-voltage field of the distribution network has increased significantly. After the distributed photovoltaic system is connected to the low-voltage distribution network, it not only reshapes the load characteristics of this area, but also exacerbates the instantaneous fluctuation of the load, thus increasing the complexity of distribution network scheduling and management. At the same time, with the rapid growth of electricity consumption, especially the surge in electricity demand during peak hours, the power grid needs to more accurately predict and dispatch power resources to ensure the balance between supply and demand. Once the prediction is inaccurate or the dispatch is untimely, it may lead to power shortages or surpluses, affecting the stable operation of the power grid. In order to effectively alleviate the challenges brought by the uncertainty of photovoltaic output to the power grid operation, it is particularly important to seek a method for predicting the net load after the grid connection of a high proportion of distributed photovoltaic power.

[0003] In recent years, deep learning methods have become the core technology for constructing high-precision net load prediction models due to their excellent feature extraction capabilities, efficient computing performance, and excellent generalization capabilities. Models involving deep learning can be roughly divided into two categories.

[0004] The first category is traditional neural network-based models. Through different improvement methods, such as weather condition classification, two-stage decomposition, Bayesian optimization, and wavelet transform, the performance of these models is optimized, especially in dealing with specific influencing factors and feature extraction.

[0005] The second category is models based on the Transformer architecture. Through multi-scale feature extraction and attention mechanisms, these models can effectively capture complex patterns and relationships, significantly improving the prediction accuracy. For example, the PSGformer model can accurately capture and separate various unique components that affect the net load prediction accuracy, and focus on these key attributes to improve the accuracy and reliability of the prediction.

[0006] Existing distributed photovoltaic short-term net load prediction algorithms have some limitations, especially in using the channel attention mechanism to enhance the feature representation ability. This deficiency may affect the prediction accuracy of the model because it reduces the ability to capture key features and may limit the generalization ability of the model on new data. In addition, existing algorithms also have limitations in considering the comprehensive impact of different influencing factors on the net load, which will further affect the prediction performance under specific conditions. Summary of the Invention

[0007] The present invention proposes a method for predicting the net load of a photovoltaic substation based on symplectic geometric mode decomposition sample entropy and squeeze-and-excitation networks. First, a net load signal decomposition model based on fluctuation complexity partitioning is established. Through symplectic geometric mode decomposition and sample entropy partitioning, multiple stationary and complex symplectic geometric components (SGCs) with significant differences are obtained. Then, a screening model for strongly correlated predictors of SGCs based on mutual information is established to determine the optimal prediction inputs for different SGCs. Finally, a combined prediction model for the net load of SGCs based on squeeze-and-excitation networks is established. For different SGCs and predictors, and by combining the prediction results, high-precision net load prediction is achieved.

[0008] Explanation of related terms:

[0009] Symplectic Geometric Mode Decomposition (SGMD): Symplectic geometric mode decomposition is an advanced signal decomposition method based on symplectic geometric theory. Through iterative symplectic transformation and screening processes, complex signals are decomposed into multiple intrinsic mode functions with different frequencies and amplitudes. SGMD is particularly suitable for processing non-linear and non-stationary signals and has the advantages of reducing mode aliasing, improving decomposition accuracy and reliability. In the field of net load data analysis, SGMD can effectively decompose complex net load signals, extract intrinsic mode functions (IMFs), and reveal the intrinsic characteristics and patterns of the data. At the same time, SGMD can also process the noise components in the data, improving the smoothness and quality of the data. In addition, this technology can also identify the trend terms in the net load data, providing important references for the long-term planning and operation scheduling of power systems.

[0010] Sample entropy: Sample entropy is a method improved on the basis of approximate entropy. Both are used to measure the probability of generating new patterns when the complexity and dimension of a time series change. The higher the probability of generating new patterns, the higher the complexity of the sequence and the corresponding entropy value. Sample entropy quantifies the complexity of a time series by comparing the similarity between adjacent subsequences of length m and calculating the conditional probability of the change in similarity between subsequences of length m + 1.

[0011] Squeeze-and-Excitation Networks (SENet): Squeeze-and-excitation networks are a network architecture that adaptively recalibrates the channel feature responses by explicitly modeling the interdependencies between convolutional features. The core of SENet is the squeeze-and-excitation block, which can be embedded into existing convolutional neural networks to enhance their sensitivity to feature channels.

[0012] The specific technical solution is as follows:

[0013] A method for predicting the net load of a power distribution area considering photovoltaic power generation, including the following processes:

[0014] Model training; specifically including the following processes:

[0015] Step S100: Collect the net load power and meteorological variable data of a certain fixed area at a fixed interval;

[0016] Step S200: Apply geometric mode decomposition to decompose the net load power by the SGMD;

[0017] Step S300: Calculate the mutual information between the net load power, its decomposed modes and meteorological variables, and construct a strong correlation predictor screening model based on mutual information to determine the best prediction inputs for different SGCs;

[0018] Step S400: Construct a combined prediction model of net load SGCs based on the squeeze-and-excitation network, and complete the model training for different SGCs and predictors;

[0019] Using the model for prediction; that is, using the model trained in the previous steps to predict the net load of the corresponding area.

[0020] Preferably, the meteorological variables in step S100 include: solar radiation, ambient temperature, wind speed, relative humidity, and timestamp.

[0021] Preferably, in step S100, the data is preprocessed, specifically including: removing outliers, filling in missing data with linear interpolation, and normalizing all numerical data.

[0022] Preferably, step S200 includes the following specific processes:

[0023] Step S210: Perform phase space reconstruction based on one-dimensional net load power to construct a trajectory matrix, and the formula is as follows:

[0024]

[0025] where x(n) = {x 1 , x 2 ,..., x n} is the one-dimensional net load power, d is the embedding dimension, and λ is the delay time;

[0026] Step S220: Construct a Hamiltonian matrix M through symplectic geometric matrix transformation, and the formula is as follows:

[0027]

[0028] Step S230: Calculate the symplectic orthogonal matrix Q, with the formula as follows:

[0029]

[0030] where B is an upper triangular matrix and N is the sequence length;

[0031] Step S240: The eigenvalues λ 1 , λ 2 ,..., λ d , with the formula as follows:

[0032]

[0033] Preferably, the step S300 includes the following specific process: Calculate the mutual information between the net load power, its decomposition modes, and meteorological variables.

[0034] Preferably, the step S400 includes the following specific process:

[0035] Step S410: The input is the time series features extracted in S300, and the long short-term memory network LSTM is used to process the input time series data;

[0036] Step S420: Global average pooling Global pooling; Squeeze the feature map of each channel into a scalar, representing the global feature response of that channel;

[0037] Step S430: Fully connected layer FC; Used for dimensionality reduction to reduce computational complexity and enhance the non-linear expression ability;

[0038] Step S440: ReLU activation function ReLU; The activation function is applied to the output of the fully connected layer to introduce non-linearity, enabling the network to learn more complex patterns;

[0039] Step S450: Fully connected layer FC; Used for dimensionality increase to match the original number of channels;

[0040] Step S460: Sigmoid activation function Sigmoid; Compress the output of the second fully connected layer between 0 and 1 to generate the weights of each feature channel;

[0041] Step S470: Feature scaling Scale; Multiply the weights of each channel obtained from the excitation operation by the corresponding channel of the original feature map to achieve feature recalibration.

[0042] When analyzing the net load with non - linear and non - stationary characteristics, the present invention uses symplectic geometric similarity transformation for adaptive decomposition and reconstruction to extract single - component signals, effectively eliminating noise interference while keeping the original time series unchanged. At the same time, by calculating the mutual information between different modes and meteorological factors, a strong - correlation predictor screening model based on mutual information, SGCs, is established to accurately select the key prediction factors affecting the net load. The squeeze - and - excitation network is introduced to enhance the model's attention to important features, enhancing the model's generalization ability, and ultimately improving the accuracy and stability of the prediction. Description of the Drawings

[0043] Figure 1 It is the original net load power signal diagram in the embodiment; the abscissa represents the time parameter, with the unit of "hour", and the ordinate represents the net load power, with the unit of "megawatt".

[0044] Figure 2 It is the signal diagram of Mode 1 decomposed by symplectic geometric decomposition in the embodiment; the abscissa represents the time parameter, with the unit of "hour", and the ordinate represents the net load power, with the unit of "megawatt"; Mode 1 shows the most significant periodic fluctuations, with a clear daily variation, reflecting the daily cycle of photovoltaic generation.

[0045] Figure 3 It is the signal diagram of Mode 2 decomposed by symplectic geometric decomposition in the embodiment; the abscissa represents the time parameter, with the unit of "hour", and the ordinate represents the net load power, with the unit of "megawatt"; Mode 2 shows a lower degree of fluctuation but retains periodicity, which may correspond to medium - term weather patterns.

[0046] Figure 4 It is the signal diagram of Mode 3 decomposed by symplectic geometric decomposition in the embodiment; the abscissa represents the time parameter, with the unit of "hour", and the ordinate represents the net load power, with the unit of "megawatt".

[0047] Figure 5 It is the signal diagram of Mode 4 decomposed by symplectic geometric decomposition in the embodiment; the abscissa represents the time parameter, with the unit of "hour", and the ordinate represents the net load power, with the unit of "megawatt".

[0048] Modes 3 and 4 show a significant increase in volatility, highlighting higher - frequency fluctuations and noise, which may indicate irregular changes in the data, such as rapid changes in solar radiation due to cloud cover.

[0049] Figure 6It is a schematic diagram of the SGCs complexity analysis results in the embodiment. The entropy of the original net load power is relatively low, indicating low information complexity. As the number of modes increases, its entropy also increases, and mode 3 shows the highest value. This indicates that mode 3 captures more detailed information and noise that other modes fail to explain. These complexity metrics help clarify the relevance of each mode in the prediction model.

[0050] Figure 7 It is a schematic diagram of the mutual information calculation results in the embodiment. The mutual information between the original net load power, its decomposed modes, and key meteorological variables - radiation, temperature, wind speed, humidity, and timestamp; stronger mutual information indicates a closer relationship between the variables and the net load or its modes, making these variables crucial for accurate prediction. In the figure, radiation has the greatest impact on modes 1 and 2, which is consistent with its observed daily cycle behavior. In addition, wind speed and humidity play a greater role in modes 3 and 4, indicating that these modes are more sensitive to short-term weather changes.

[0051] Figure 8 It is a schematic diagram of the squeeze-and-excitation network structure in the embodiment.

[0052] Figure 9 It is a schematic diagram comparing the net load prediction results of various models under clear sky conditions in the embodiment; the performance of the model of the present invention and other prediction models is compared under these two weather conditions. Under clear sky conditions, the prediction results of all models are relatively close to the actual load curve, showing low errors and better prediction consistency. This may be because the predictability of solar radiation is relatively high, thus enhancing the overall model performance.

[0053] Figure 10 It is a schematic diagram comparing the net load prediction results of various models under cloudy conditions in the embodiment; under cloudy conditions, the prediction errors generally increase, especially when the cloud cover changes cause significant fluctuations in the photovoltaic output. Nevertheless, the model of the present invention still maintains a relatively low error level, highlighting its adaptability to net load fluctuations under unstable weather conditions. Detailed implementation manner

[0054] A method for predicting the net load of a distribution area considering photovoltaic power generation includes the following processes:

[0055] Model training; specifically includes the following processes:

[0056] Step S100: The data set contains the net load power and meteorological data sampled every 15 minutes in a distributed photovoltaic area in southern China from 2019 to 2021. The meteorological data includes: solar radiation, ambient temperature, wind speed, relative humidity, and timestamp; it is used to analyze the influence of sunshine, seasons, and holidays.

[0057] During data preprocessing, outliers were removed, missing data were filled using linear interpolation, and all numerical data were standardized. The training set consisted of data from 2019 and 2020, accounting for 70% of the total dataset. The validation set was used for model selection and hyperparameter tuning, including the remaining data from 2020, accounting for 20%. The test set contained data from 2021, accounting for 10% of the total dataset, and was used for the final evaluation of the model performance.

[0058] Step S200: Apply geometric mode decomposition to the SGMD to decompose the net load power into modes. Step S210: Based on the one-dimensional net load power, perform phase space reconstruction to construct a trajectory matrix, with the formula as follows:

[0059]

[0060] where \(x(n)=\{x 1 ,x 2 ,\cdots,x n \}\) is the one-dimensional net load power, \(d\) is the embedding dimension, and \(\lambda\) is the delay time.

[0061] Step S220: Through symplectic geometric matrix transformation, construct the Hamiltonian matrix \(M\), with the formula as follows:

[0062]

[0063] Step S230: Calculate the symplectic orthogonal matrix \(Q\), with the formula as follows:

[0064]

[0065] where \(B\) is an upper triangular matrix and \(N\) is the sequence length.

[0066] Step S240: The eigenvalues \(\lambda 1 ,\lambda 2 ,\cdots,\lambda d of matrix \(A\), with the formula as follows:

[0067]

[0068] Figures 1 - 5 shows the original net load power and the four main modal signals decomposed therefrom.

[0069] Step S250: Calculate the sample entropy; the sample entropy is a complexity metric method improved on the basis of the approximate entropy. They are both used to measure the probability of generating new patterns when the time series changes in complexity and dimension. The higher the probability of generating new patterns, the higher the complexity of the sequence and the corresponding entropy value. The sample entropy quantifies the complexity of the time series by comparing the similarity between adjacent subsequences of length m and calculating the conditional probability of the similarity change between subsequences of length m+1. The calculation steps of the sample entropy are as follows:

[0070] Step S251: Set the parameters m and r, where m is the length of the subsequence and r is the similarity threshold;

[0071] Step S252: Calculate the similarity between each subsequence of length m in the time series and all other subsequences of length m;

[0072] Step S253: Count the number of subsequence pairs with similarity less than r and calculate its proportion among all subsequence pairs, denoted as B m (r); The calculation formula is as follows:

[0073]

[0074] Step S254: For subsequences of length m+1, calculate the proportion of subsequence pairs with similarity less than r, denoted as A m (r); The calculation formula is as follows:

[0075]

[0076] Step S255: The sample entropy reflects the probability of the time series generating new patterns;

[0077] where N is the length of the time series, A i is the number of sequence pairs with distance less than r when the dimension increases to m+1, and B i is the number of sequence pairs with distance less than r when the dimension is m;

[0078] Step S300: Calculate the mutual information between the net load power, its decomposition modes and meteorological variables, and construct a strong correlation predictor screening model based on mutual information to determine the best prediction inputs for different SGCs; including the following specific processes:

[0079] This model calculates the mutual information between each SGC and potential predictors, quantifies the statistical dependence between them, and accurately identifies the factor combinations that contribute the most to the prediction of specific SGCs; calculate the mutual information between the net load power, its decomposition modes and meteorological variables;

[0080] Stronger mutual information indicates a closer relationship between variables and the net load or its modes, making these variables crucial for accurate prediction. For example, radiation has the greatest impact on modes 1 and 2, which is consistent with its observed daily periodic behavior. In addition, wind speed and humidity play a greater role in modes 3 and 4, indicating that these modes are more sensitive to short-term weather changes. As Figure 7 shown;

[0081] Step S400: Based on the squeeze-and-excitation network, construct a combined prediction model for net load SGCs, and complete model training for different SGCs and prediction factors. The specific process of step S400 is as follows:

[0082] Step S410: The input is the time series features extracted in S300, and the long short-term memory network (LSTM) is used to process the input time series data;

[0083] Step S420: Global average pooling (Global pooling); squeeze the feature map of each channel into a scalar, representing the global feature response of that channel; each channel is squeezed into a single numerical value to achieve global perception of the channel.

[0084] Step S430: Fully connected layer (FC); used for dimensionality reduction to reduce computational complexity and enhance non-linear expression ability; through this fully connected layer, the network can capture the correlation between features and extract the most important feature representations through dimensionality reduction.

[0085] Step S440: ReLU activation function (ReLU); the activation function is applied to the output of the fully connected layer to introduce non-linearity, enabling the network to learn more complex patterns; the ReLU function sets all negative values to zero and keeps positive values unchanged.

[0086] Step S450: Fully connected layer (FC); used for dimensionality increase to match the original number of channels;

[0087] Through this fully connected layer, the network can learn the weights of each channel, which represent the importance of that channel for the current task.

[0088] Step S460: Sigmoid activation function (Sigmoid); compress the output of the second fully connected layer between 0 and 1 to generate the weights of each feature channel; these weights represent the importance of each feature channel for the final prediction.

[0089] Step S470: Feature scaling (Scale); multiply the weights of each channel obtained from the excitation operation by the corresponding channel of the original feature map to achieve re-calibration of the features; essentially, it is a weighted operation on the original feature map, and the weights are learned by the SE block.

[0090] The weighted feature map will have stronger representation ability because it emphasizes important channels and suppresses less useful channels.

[0091] Comparison of prediction accuracies of different models:

[0092] In this embodiment, real data is used to validate the proposed photovoltaic net load prediction method, and its performance is compared with several existing models. The accuracy comparison table shows the comparison between the model of the present invention and other commonly used prediction models. As shown in the following table.

[0093] Model MAPE / % RMSE / MW MAE / MW Proposed 6.79 17.82 12.11 XGBOOST 7.69 20.43 15.42 Bi - LSTM 9.35 20.06 17.52 Transformer 11.03 21.93 19.77 TCN 9.54 18.68 15.81 Nbeats 9.56 18.57 14.34

[0094] In terms of MAPE, the model of the present invention exceeds all other models with a value of 6.79%, which is 2.64% lower than the traditional method. The MAPE values of the XGBoost and TCN models are 7.69% and 9.54% respectively. Although relatively low, there is still an obvious gap compared with the model of the present invention. The MAPE values of Bi-LSTM and N-BEATS are 9.35% and 9.56% respectively, while the MAPE value of the Transformer model is the highest at 11.03%, indicating that it performs the worst on the dataset used in the present invention.

[0095] For the RMSE metric, the model of the present invention again shows a relatively low error value of 17.82 MW, indicating that its prediction fluctuations are closely aligned with the actual data. In contrast, the RMSE of XGBoost is 20.43 MW, while the Transformer model performs the worst with an RMSE of 21.93 MW. The RMSE values of TCN and N-BEATS are 18.68 MW and 18.57 MW respectively, reflecting reasonable error control.

[0096] Regarding the MAE metric, the model of the present invention leads again with a value of 12.11 MW, further confirming its robustness in maintaining a low average error. Other models, including XGBoost, TCN, and N-BEATS, have MAE values of 15.42 MW, 15.81 MW, and 14.34 MW respectively, all higher than the model of the present invention. The MAE values of Bi-LSTM and Transformer are 17.52 MW and 19.77 MW respectively, indicating that these models have greater difficulty in maintaining a low average error.

[0097] In the present invention, the original net load is processed by using symplectic geometric mode decomposition and sample entropy to obtain symplectic geometric components (SGCs) that are stationary and have significant features. Subsequently, the mutual information between the original net load and its decomposed modes and key meteorological variables such as radiation, temperature, wind speed, humidity, and time is calculated. Based on the selected strong mutual information, a squeeze-and-excitation network is used for prediction. For the net load data of a specific photovoltaic area, the following conclusions are established: the model of the present invention is superior to other benchmark models in terms of prediction accuracy and generalization ability. The experimental results show that the model of the present invention is superior to other benchmark models in evaluation indexes such as MAPE, RMSE, and MAE, and the hypothesis holds, which proves the effectiveness of the prediction method for the short-term net load prediction of high-proportion distributed photovoltaic power.

Claims

1. A method for predicting net load of a photovoltaic area, characterized in that: The process includes the following: Model training; specifically includes the following processes: Step S100: Collecting net load power and meteorological variable data in a fixed area and at a fixed interval; Step S200: Applying geometric mode decomposition (SGMD) to perform modal decomposition on the net load power; Step S300: Calculate the mutual information between the net load power, its decomposed modes and meteorological variables, and construct a SGCs strong correlation prediction factor screening model based on the mutual information to determine the best prediction inputs for different SGCs; Step S400: constructing a net load SGCs combined prediction model based on the squeeze incentive network, and completing model training for different SGCs and prediction factors; Use the model to make predictions; That is: use the model trained in the previous steps to predict the net load of the corresponding area.

2. According to claim 1, a method for predicting net load of a photovoltaic area is characterized in that: The meteorological variables in step S100 include: solar radiation, ambient temperature, wind speed, relative humidity, and timestamp.

3. According to claim 2, a method for predicting net load of a photovoltaic area is characterized in that: In step S100, the data is preprocessed, specifically including: Outliers were removed, missing data were filled using linear interpolation, and all numerical data were normalized.

4. According to claim 3, a method for predicting net load of a photovoltaic area is characterized in that: The step S200 includes the following specific processes: Step S210: Reconstruct the phase space based on the one-dimensional net load power to construct a trajectory matrix, the formula is as follows: Where x(n)={x1,x2,...,x n } is the one-dimensional net load power, d is the embedding dimension, and λ is the delay time; Step S220: construct the Hamiltonian matrix M through symplectic geometric matrix transformation, the formula is as follows: Step S230: Calculate the symplectic orthogonal matrix Q, the formula is as follows: Where B is an upper triangular matrix and N is the sequence length; Step S240: eigenvalues ​​λ1,λ2,...,λ of matrix A d , the formula is as follows:

5. According to claim 3, a method for predicting net load of a photovoltaic power station is characterized in that: The step S300 includes the following specific processes: The mutual information between the net load power, its decomposed modes and meteorological variables is calculated.

6. According to claim 3, a method for predicting net load of a substation considering photovoltaic power generation is characterized in that: The step S400 includes the following specific processes: Step S410: The input is the time series features extracted in S300, and a long short-term memory network (LSTM) is used to process the input time series data; Step S420: Global average pooling: squeezing the feature map of each channel into a scalar, representing the global feature response of the channel; Step S430: Fully connected layer (FC); used for dimensionality reduction to reduce computational complexity and enhance nonlinear expression capability; Step S440: ReLU activation function (ReLU); the activation function is applied to the fully connected layer output, introducing nonlinearity so that the network can learn more complex patterns; Step S450: Fully connected layer (FC); used to increase the dimension to match the original number of channels; Step S460: Sigmoid activation function (Sigmoid); compress the output of the second fully connected layer to between 0 and 1, and generate the weight of each feature channel; Step S470: Feature scaling (Scale); multiply the weight of each channel obtained by the excitation operation by the corresponding channel of the original feature map, Enables recalibration of features.

Images