Power load prediction method based on double-layer decomposition and deep learning integration

Abstract

The invention provides a power load prediction method based on double-layer decomposition and deep learning integration, belongs to the field of power system scheduling and energy prediction, and is suitable for hour-level and day-level multi-scale load prediction. The method aims to overcome the defects that an existing method is not thorough in decomposition, large in noise interference, weak in time sequence feature capture and simple in integration strategy, and prediction precision and robustness are improved. Comprising the following steps: acquiring original power load containing time sequence characteristics and associated influence factor data, and preprocessing; data is decomposed through CEEMDAN, high and low frequency components are distinguished and reconstructed, and high-frequency signals are denoised through a wavelet threshold method; carrying out secondary decomposition on the fusion signal by the VMD to obtain a stationary target IMF; a BiTCN (BiTCN-BiGRU)-Attention model is used for respectively predicting each IMF (Inertial Measurement Function); and based on a variance contribution degree weighted accumulation prediction value, a final result can provide data support for power generation planning, power grid dispatching and energy storage optimization of a power system.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and energy forecasting technology, specifically to a power load forecasting method based on two-layer signal decomposition and deep learning integration. It is applicable to power load forecasting at multiple scales, such as hourly and daily levels, and can provide high-precision data support for power system generation planning, grid dispatching, and energy storage optimization. Background Technology

[0002] Electricity load forecasting is a core component of power system operation and management, and its accuracy directly affects the safety, economy, and reliability of the power grid. However, the original power load data is affected by many factors such as residents' electricity consumption habits, industrial production fluctuations, changes in meteorological conditions and adjustments in electricity price policies, and exhibits strong nonlinearity, multi-frequency fluctuations and noise interference. This leads to the following key problems in traditional prediction methods: (1) Limited data decomposition effect: Single signal decomposition methods are difficult to process high-frequency noise and low-frequency trends in load data at the same time. EMD is prone to "modal aliasing", resulting in insufficient stability of the decomposed subsequences. VMD is sensitive to strong noise in the original data, and direct decomposition is prone to introducing pseudo features. (2) Incomplete noise processing: High-frequency components in power load data often contain random noise. Traditional filtering methods are prone to smoothing useful signals, resulting in feature loss. (3) Weak ability to capture time series features: Traditional machine learning models are difficult to capture the long-term dependence and multi-scale correlation of load data. Single deep learning models either cannot take into account forward and backward time series information or do not pay enough attention to the features of key time steps, resulting in limited prediction accuracy. (4) Simple integration strategy: Existing multi-component prediction integration methods mostly adopt equal weight accumulation, without considering the differences in contribution of different components to the final load, and cannot give full play to the prediction value of each component.

[0003] To address the aforementioned issues, an integrated approach of "decomposition-denoising-re-decomposition-accurate prediction-intelligent integration" needs to be designed. This approach combines multi-level signal processing with deep learning to improve the accuracy and robustness of power load forecasting in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing power load forecasting methods, such as "incomplete decomposition, large noise interference, weak feature capture, and poor integration effect", and to provide a power load forecasting method based on two-layer decomposition and deep learning integration, so as to achieve high-precision forecasting of highly nonlinear and noisy power load data.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A power load forecasting method based on bi-level decomposition and deep learning integration includes:

[0007] Electricity load data acquisition: Collect raw electricity load data with time-series characteristics. The raw electricity load data shall include at least hourly electricity load values ​​and be synchronously correlated with influencing factors such as dry-bulb temperature, dew point temperature, humidity, and electricity price.

[0008] Decomposition and High / Low Frequency Reconstruction: The CEEMDAN algorithm is used to decompose the original power load data to obtain several intrinsic mode functions (IMFs). The reconstruction algorithm distinguishes the high-frequency components and low-frequency components in the IMFs. The sum of j consecutive IMFs is used as the reconstruction index j. The mean of each index j is calculated and g test is performed to screen out the stable high-frequency component set and low-frequency component set. The two sets are then reconstructed by accumulation to obtain the initial high-frequency reconstruction signal and the initial low-frequency reconstruction signal.

[0009] High-frequency component denoising: The initial high-frequency reconstructed signal is denoised using wavelet thresholding, specifically including: Wavelet decomposition: N-level wavelet decomposition of the initial high-frequency reconstructed signal is performed using wavelet basis functions to obtain N-level high-frequency detail coefficients and 1-level low-frequency approximation coefficients; Thresholding: The N-level high-frequency detail coefficients are truncated using soft or hard thresholding functions to remove coefficients corresponding to noise; Signal reconstruction: The processed high-frequency detail coefficients and the low-frequency approximation coefficients are subjected to inverse wavelet decomposition to obtain the denoised high-frequency signal.

[0010] The second feature extraction is as follows:

[0011] The fused signal X fusion =L0+H1, input VMD algorithm,

[0012] Furthermore, the parameters were set as follows: number of modes K = 5, penalty factor α = 2000, noise tolerance = 1e-7; the variational optimization problem was solved using the ADMM algorithm to minimize the sum of the bandwidths of each IMF, and finally 5 target IMFs (IMF_a~IMF_e) were obtained. The center frequencies of each IMF covered 0.01~0.1Hz, 0.1~0.5Hz, 0.5~1Hz, 1~2Hz, and 2~5Hz, respectively. The stationarity was improved by more than 40% compared with the original data, and then the sample entropy index was used for verification.

[0013] The deep learning model predicts as follows:

[0014] A BiTCN-BiGRU-Attention prediction model is constructed to predict the load of each target IMF. Further, in the input layer, the time-series data of a single target IMF is divided into samples using a sliding window method of "predicting the load of the next hour from the data of the previous 24 hours" to obtain the input features.

[0015] Simultaneously incorporating influencing factors such as temperature and humidity, expanding the input dimensions to...

[0016]

[0017] Furthermore, the BiTCN module contains two causal convolutional layers with a kernel size of 3 and a total of 64 kernels. Each layer is activated by ReLU and connected to residuals, outputting multi-scale features.

[0018] BiGRU module: Contains 1 layer forward GRU and 1 layer backward GRU, with 128 hidden units, and supports F... TCN Perform bidirectional time series learning and output

[0019]

[0020] Furthermore, the Attention module: uses a linear transformation to convert F... GRU The mapping is done using query vector Q, key vector K, and value vector V, and attention weights are calculated.

[0021] Output weighted features

[0022] Output layer: Outputs the predicted value through a fully connected layer;

[0023] Furthermore, for model training: MSE was used as the loss function, AdamW optimizer was used, 100 training epochs were conducted, batch size was 32, and an early stopping strategy was adopted to prevent overfitting.

[0024] The prediction results are integrated and fused as follows:

[0025] Furthermore, the variances of the five target IMFs were calculated. Weighting coefficient

[0026] The final prediction result is obtained by summing.

[0027]

[0028] The power load forecasting method based on bi-layer decomposition and deep learning integration provided by this invention has the following beneficial effects:

[0029] More thorough decomposition and higher data quality: Through CEEMD-VMD two-layer decomposition, high and low frequency components are first separated and denoised, and then multi-frequency features are extracted a second time, reducing sample entropy by 40% to 60% and effectively solving the problems of "modal mixing" and "noise interference"; More accurate feature capture: The BiTCN-BiGRU-Attention model integrates multi-scale dependence, bidirectional time series and key feature focusing, improving the ability to capture nonlinear trends in load data by more than 30%; More intelligent integration: The weighted strategy based on variance contribution fully leverages the predictive value of each component, reducing MAE by 15% to 25% compared to equal-weighted integration; Stronger robustness: Validated on the Australian electricity load dataset, MAE = 28.5MW, RMSE = 36.2MW, MAPE = 1.2%, improving accuracy by 20% to 40% compared to ARIMA and single GRU models, and is suitable for complex meteorological and power consumption fluctuation scenarios. Attached Figure Description

[0030] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart of the power load forecasting method of the present invention;

[0032] Figure 2 This is a diagram showing the VMD secondary decomposition results of the power load data in this invention;

[0033] Figure 3 This is a comparison chart of the power load prediction results of this invention. Detailed Implementation

[0034] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0035] The following are specific embodiments of the present invention:

[0036] Step 1, Data Preprocessing:

[0037] The 12 missing records in the dataset were filled using linear interpolation. The formula is as follows:

[0038]

[0039] In the formula, x i x represents the load value at the missing time.i-1 For the missing moment, x i-1 x i+1 These represent the effective load values ​​adjacent to the missing time point.

[0040] Using the 3σ criterion (values ​​exceeding the mean ± 3 times the standard deviation are considered abnormal), 8 extreme abnormal data points were removed (such as load values ​​> 12000MW or < 5000MW, caused by sudden failures of industrial equipment or data transmission errors).

[0041] Step 2, CEEMDAN decomposition and high / low frequency reconstruction:

[0042] Gaussian white noise with an amplitude of 0.1 times the standard deviation of the original data was added to the standardized raw power load data. Through iterative decomposition using the EMD algorithm, eight intrinsic mode functions (IMFs) and one residual trend term were finally obtained. The mean values ​​for “IMF1”, “IMF1+IMF2”, …, “IMF1+…+IMF8” were calculated respectively. with standard deviation s j The high-frequency and low-frequency components are determined by the g-test statistic. The g-test formula is:

[0043]

[0044] In the formula, n = 8760 is the total number of samples, when g j When the value is greater than 1.96, the corresponding IMF is classified as a high-frequency component; otherwise, it is classified as a low-frequency component. The high-frequency component set and the low-frequency component set are accumulated and reconstructed respectively to obtain the initial high-frequency reconstructed signal H0 and the initial low-frequency reconstructed signal L0, which are prepared for subsequent denoising processing.

[0045] Step 3, High-frequency component noise reduction:

[0046] The db4 wavelet is used to perform a 4-level wavelet decomposition on the initial high-frequency reconstructed signal H0, obtaining 4 levels of high-frequency detail coefficients and 1 level of low-frequency approximation coefficients. A soft thresholding function is used to truncate the high-frequency detail coefficients to avoid signal abrupt changes caused by hard thresholding. The formula for the soft thresholding function is:

[0047] w λ (x)=sign(x)×max(|x|-λ,0) (3)

[0048] In the formula, λ is the threshold, calculated using the Birgé-Massart strategy to be λ = 18.6, which is the standard deviation of d1. The processed high-frequency detail coefficients and low-frequency approximation coefficients are subjected to wavelet inverse decomposition to obtain the denoised high-frequency signal H1, which is then fused with the initial low-frequency reconstructed signal L0 to form X. fusion =L0+H1.

[0049] Step 4, Second Feature Extraction:

[0050] The fused signal X fusion The VMD algorithm is input, and more stable load characteristics are extracted through the variational framework. The variational optimization problem is solved by the alternating direction multiplier method to minimize the sum of the bandwidths of each IMF, and finally five target IMFs are obtained.

[0051] Step 5, BiTCN-BiGRU-Attention model training:

[0052] For each target IMF, an independent prediction model is constructed. Then, a sliding window strategy of "predicting the load of the next hour from the data of the previous 24 hours" is adopted. The input feature dimension is 24×6, and the output is the power load value of the 25th hour. The input layer receives a 24×6 feature matrix, which is converted into a 24×6 time-series vector through a reshape layer. The BiTCN module contains two causal convolutional layers with a kernel size of 3 and 64 kernels per layer. After processing with ReLU activation function and residual connection, it captures the multi-scale temporal dependencies of the load data. The BiGRU module contains one forward GRU layer and one backward GRU layer with 128 hidden units, which learns the forward and backward correlation information of the time-series data. The Attention module adopts an additive attention mechanism to assign weights to the 256-dimensional bidirectional features output by BiGRU to highlight the influence of key time steps. The weight calculation formula is as follows:

[0053]

[0054] In the formula, h t Let u be the output of BiGRU at time t, and u be the attention context vector.

[0055] score(·) is a linear scoring function; Output layer: Outputs the predicted value of a single IMF through a fully connected layer.

[0056] Step Six: Integration and Fusion of Prediction Results:

[0057] Weight Calculation: The integration weight is determined based on the variance contribution of each target IMF. The larger the variance of an IMF, the more significant its impact on the final load. The weighting formula is as follows:

[0058]

[0059] In the formula, Let V be the variance of the k-th IMF;

[0060] The electricity load forecast is a weighted sum of the IMF forecasts, using the following formula:

[0061]

[0062] In the formula, y is the final predicted value. This is the predicted value for the k-th IMF.

[0063] Combination Figure 1 Power load forecasting method technical flowchart:

[0064] As shown in the figure, the original data is first decomposed into high and low frequencies using CEEMDAN. The high-frequency components are then denoised by wavelet thresholding and fused with the low-frequency components before being decomposed again by VMD to obtain a stable target IMF. Next, the BiTCN-BiGRU-Attention model is used to predict each IMF separately. Finally, the final power load prediction value is output through variance-weighted integration. The links are closely connected, which effectively solves the problems of incomplete decomposition and low accuracy in traditional prediction.

[0065] Combination Figure 2 VMD secondary decomposition result diagram of power load data:

[0066] As shown in the figure, the time-domain waveforms of the five target IMFs (IMF_a to IMF_e) obtained by VMD secondary decomposition were analyzed. The frequency layering of each component is clear: from low-frequency trends (such as IMF_a, 0.05Hz) to high-frequency details (such as IMF_e, 2.5Hz), there is no mode mixing and the stability is significantly improved, providing a clear feature basis for subsequent accurate prediction of components.

[0067] Combination Figure 3 Comparison chart of power load forecast results:

[0068] As shown in the figure, a time-series comparison is made between the predicted values ​​of the method of this invention and the single BiGRU method, and the actual power load values. It is evident that the predicted curve of this invention has the highest degree of overlap with the actual curve, especially during critical periods such as load peaks and extreme weather events, where the deviation is far smaller than that of the other method, directly demonstrating its superior prediction accuracy.

[0069] The above-described embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any simple changes or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the scope of the technology disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A power load forecasting method based on bi-level decomposition and deep learning integration, characterized in that, include: (1) Power load data acquisition: Collect raw power load data with time-series characteristics. The raw power load data shall include at least hourly power load values ​​and be synchronously correlated with data on influencing factors such as dry bulb temperature, dew point temperature, humidity, and electricity price. (2) Decomposition and High- and Low-Frequency Reconstruction: The original power load data is decomposed using the CEEMDAN algorithm to obtain several intrinsic mode functions (IMFs); the high-frequency components and low-frequency components in the IMFs are distinguished by the reconstruction algorithm, and the sum of j consecutive IMFs is used as the reconstruction index j. The mean of each index j is calculated and g test is performed to screen out the stable high-frequency component set and low-frequency component set. The two sets are reconstructed by accumulation to obtain the initial high-frequency reconstruction signal and the initial low-frequency reconstruction signal. (3) High-frequency component denoising: The initial high-frequency reconstructed signal obtained in step (2) is denoised using the wavelet thresholding method, specifically including: Wavelet decomposition: Select wavelet basis functions to perform N-level wavelet decomposition on the initial high-frequency reconstructed signal to obtain N-level high-frequency detail coefficients and 1-level low-frequency approximation coefficients; Thresholding: The N layers of high-frequency detail coefficients are truncated using a soft or hard thresholding function to remove coefficients corresponding to noise. Signal reconstruction: The processed high-frequency detail coefficients and the low-frequency approximation coefficients are subjected to wavelet inverse decomposition to obtain the denoised high-frequency signal; (4) Second feature extraction: The VMD algorithm is used to perform a second decomposition on the fused signal composed of "the initial low-frequency reconstructed signal in step (2) + the denoised high-frequency signal in step (3)". The VMD modality number K, penalty factor α and noise tolerance are set to obtain K target IMFs with enhanced stability and different frequency features. (5) Deep learning model prediction: Construct a BiTCN-BiGRU-Attention prediction model to predict each target IMF obtained in step (4). The model includes: Input layer: Receives time-series data and related influencing factor data of a single target IMF, and performs standardized preprocessing; The BiTCN module contains multiple causal convolutional layers and residual connections. It uses convolutional kernels of different sizes to capture multi-scale dependencies of the input temporal data and outputs multi-scale feature vectors. BiGRU module: Composed of forward GRU layer and backward GRU layer, it receives feature vectors output by BiTCN module, learns the forward and backward dependencies of time series data, and outputs bidirectional time series features; Attention module: It adopts additive attention or multiplicative attention mechanism to assign weights to the bidirectional temporal features output by the BiGRU module, highlight the features of key time steps, and output a weighted feature vector. Output layer: The weighted feature vector is mapped through a fully connected layer to output the predicted value of the IMF for a single target. (6) Prediction result integration: The predicted values ​​of the K target IMFs obtained in step (5) are integrated using a weighted accumulation strategy. The weight coefficients are determined based on the variance contribution of each target IMF, and the final power load prediction result is obtained after accumulation.

2. The power load forecasting method based on bi-layer decomposition and deep learning integration according to claim 1, characterized in that, In step (2), the specific implementation of the CEEMDAN algorithm includes: adding Gaussian white noise to the original power load data, performing multi-round decomposition using the EMD algorithm, calculating the mean of the residual components of each round of decomposition, and obtaining several final IMFs; the statistic for the g-test is: in Let s be the mean of index j. j Let g be the standard deviation of index j, and n be the sample size. j If the mean is greater than 1.96, it is considered to be significantly different from 0, and the corresponding IMF is classified as a high-frequency component; otherwise, it is classified as a low-frequency component.

3. The power load forecasting method based on bi-layer decomposition and deep learning integration according to claim 1, characterized in that, In step (3), the wavelet basis function is selected from db4, sym8, or coif5 wavelets. The number of wavelet decomposition levels N is determined according to the sampling frequency of the original power load data. When the sampling frequency is on the hourly level, N = 3 to 5 levels. The threshold function adopts a soft threshold function as follows: w λ (x)=sign(x)max(|x|-λ,0), Where λ is the threshold, the specific value of which can be determined using the Birgé-Massart strategy.

4. The power load forecasting method based on bi-layer decomposition and deep learning integration according to claim 1, characterized in that, In step (4), the parameters of the VMD algorithm are set as follows: number of modes K = 3 to 8, penalty factor α = 1000 to 3000, noise tolerance = 1e-7, and upper limit of iteration number = 1000. The decomposition process solves the variational problem by alternating direction multiplier method so that the center frequencies of each target IMF do not overlap.

5. The power load forecasting method based on bi-layer decomposition and deep learning integration according to claim 1, characterized in that, In step (5), the causal convolutional layer of the BiTCN module adopts 1D convolution with a kernel size of 3 to 7, and each convolution is processed by the ReLU activation function; the BiGRU module contains 1 to 3 hidden layers with 64 to 256 hidden units in each layer; the weighting coefficients of the Attention module are calculated by linear transformation of the BiGRU output features and the softmax function.

6. The power load forecasting method according to claim 1, characterized in that, In step (6), the weight coefficients of each target IMF are calculated using the variance contribution: Let the variance of the k-th target IMF be... Then the weighting coefficient Final prediction results in Let be the predicted value of the k-th target IMF.

Images