A load prediction method and system fusing bimodal decomposition and attention mechanism

Abstract

The application provides a load prediction method and system fusing bimodule decomposition and attention mechanism, which is used for predicting future load data according to original load data. The original load data is processed through cascade processing of complete ensemble empirical mode decomposition and variational mode decomposition, high-frequency noise is removed, residual signals and corresponding multi-scale feature components are obtained, the multi-scale feature components are recombined according to time sequences to form a multi-dimensional input sample matrix, a part of the multi-dimensional input sample matrix is used to complete model training as a training set, and another part of the multi-dimensional input sample matrix is used to complete model testing as a test set, and future load data is predicted according to the trained model.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and specifically to a load forecasting method and system that integrates dual-mode decomposition and attention mechanism. Background Technology

[0002] With the continuous expansion of industrial scale and the continuous improvement of people's living standards, my country's total electricity consumption and load are also growing rapidly. As a key infrastructure supporting the normal operation of modern society, the scale and complexity of the power system are also constantly increasing. In particular, the grid connection of new energy sources with volatility and randomness, as well as new loads with high randomness, has made the balance between load supply and demand in the power system more complex and increased the uncertainty of system operation. Therefore, accelerating the construction of smart grids and improving the intelligence and informatization level of the power system has become the main direction of my country's power system development.

[0003] Power system load forecasting is a crucial component of the safe, stable operation and economically reliable dispatch of the power grid. With changes in urban population, urban economic development, natural environment, and international energy markets, power grids are becoming increasingly complex, making forecast accuracy susceptible to uncertainty and volatility. Researching high-performance load forecasting models can provide more accurate information on future load changes, offering decision-makers a more reliable basis for decision-making.

[0004] Current mainstream technologies can be divided into three categories: statistical models, such as ARIMA and SARIMA, which rely on linear assumptions and cannot handle non-stationary signals; single-stage decomposition + single-layer model, such as EMD-LSTM and VMD-SVR, but suffer from severe mode aliasing in EMD-type methods (e.g., high-frequency IMF components mixed with fluctuations of different scales) and the inability of single-stage decomposition to solve the component overlap problem; and deep learning, such as Transformer and TCN, which can capture nonlinear relationships, but are insufficient in modeling the multidimensional characteristics of load data. Summary of the Invention

[0005] This invention addresses the problems existing in the prior art by providing a load forecasting method and system that integrates dual-mode decomposition and attention mechanisms. By combining dual-mode decomposition with an improved BiLSTM load forecasting model, the complex multidimensional nonlinear characteristics of load data are decomposed through dual-mode decomposition. Targeting the temporal characteristics of the load, the combination of BiLSTM and attention mechanisms enables it to better capture the temporal characteristics of load data. Furthermore, the improved model accurately analyzes and learns the strong temporal characteristics and complex multidimensional nonlinear features of power load data, thereby achieving effective and reliable load forecasting.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of the present invention provides a load prediction method that integrates dual-mode decomposition and attention mechanism. The method is used to predict second load data from first load signal data, including: removing high-frequency noise from the first load signal by cascading processing of complete ensemble empirical mode decomposition (CEEMD) and variational mode decomposition (VMD), including: performing data decomposition on the first load signal by CEEMD to obtain a residual signal; inputting the residual signal into VMD to perform local frequency band decomposition on the residual signal to obtain multi-scale feature components.

[0007] The multi-scale feature components are reorganized according to the time series to form a multi-dimensional input sample matrix. The multi-dimensional input sample matrix includes a first sample matrix and a second sample matrix. The first sample matrix and the second sample matrix include time-aligned trend features, periodic features and transient disturbance components to represent the multi-layer feature information of the load time series signal.

[0008] The first sample matrix is ​​sequentially input into a bidirectional long short-term memory network (Bi-LSTM) for model training, including: extracting bidirectional hidden state vectors based on the potential correlation between historical information and future data modeled by the forward and backward channels in the Bi-LSTM; constructing a dynamic attention weight matrix based on the bidirectional hidden state vectors; calculating the importance score of features at each time step through an adaptive weight adjustment mechanism; and performing weighted fusion in the time dimension to obtain a global temporal feature representation, followed by iterative calculation.

[0009] The trained Bi-LSTM model is used to predict the second sample matrix to obtain the second load data.

[0010] According to an embodiment of the present invention, the step of performing data decomposition on the first load signal by CEEMD to obtain a residual signal includes: injecting different Gaussian noise sequences into the first load signal to obtain a noisy signal set; performing empirical mode decomposition on each group of signals in the noisy signal set and calculating the mean of each intrinsic mode function, and peeling off high-frequency components layer by layer to obtain a residual signal.

[0011] According to one embodiment of the present invention, the cascading of CEEMD and VMD adopts an adaptive threshold determination mechanism, using the rate of change of the spectral energy density of the residual signal as the termination condition for the decomposition depth, and dynamically determining the number of decomposition layers.

[0012] According to an embodiment of the present invention, the step of constructing a dynamic attention weight matrix based on the bidirectional hidden state vector, calculating the importance score of the feature at each time step through an adaptive weight adjustment mechanism, and performing weighted fusion in the time dimension to obtain a global temporal feature representation and performing iterative calculation includes: Step S1: Calculating the forward hidden state vector and the backward hidden state vector respectively through a dual-channel gating unit, and calculating the attention score by combining the current context information through a gating mechanism.

[0013] Step S2: Normalize the attention score to obtain the attention weight at each time step; obtain the context vector by weighted summation.

[0014] According to an embodiment of the present invention, the step of constructing a dynamic attention weight matrix based on the bidirectional hidden state vector, calculating the importance scores of features at each time step through an adaptive weight adjustment mechanism, and performing weighted fusion in the time dimension to obtain a global temporal feature representation and performing iterative calculation further includes: S3: Set the iteration number threshold and the time step prediction result error function. The time step prediction result error function is expressed as the average value of the prediction result error of all time steps within a preset time period; set the average threshold of the time step prediction result error.

[0015] S4: Generate a first test result based on the context vector and compare it with the first sample matrix. Determine whether training has ended based on the comparison result, including: if the difference between the first test result and the data in the first sample matrix is ​​less than the average threshold of the prediction result error of all time steps within a preset time period, and is greater than the iteration number threshold, then training ends.

[0016] If the difference between the first test result and the first sample matrix data is not less than the average threshold of the prediction result error of all time steps within a preset time period, or is not greater than the iteration number threshold, return to S1. The step of constructing a dynamic attention weight matrix based on the bidirectional hidden state vector, calculating the importance score of features at each time step through an adaptive weight adjustment mechanism, and performing weighted fusion in the time dimension to obtain a global temporal feature representation and performing iterative calculation further includes: Step S3: Set the iteration number threshold and the time step prediction result error function, wherein the time step prediction result error function is expressed as the average value of the prediction result errors of all time steps within the time period; set the average threshold of the time step prediction result error.

[0017] Step S4: Generate a first test result based on the context vector and compare it with the first sample matrix. Determine whether training has ended based on the comparison result, including: if the difference between the first test result and the data in the first sample matrix is ​​less than the average threshold of the prediction result error of all time steps within a preset time period, and is greater than the iteration number threshold, then training ends.

[0018] If the difference between the first test result and the first sample matrix data is not less than the average threshold of the prediction result error of all time steps within the preset time period, or is not greater than the iteration number threshold, return to S1.

[0019] According to one embodiment of the present invention, the step of using a trained Bi-LSTM model to predict the second sample matrix to obtain second load data includes: taking the second sample matrix as input, obtaining second test data according to the trained model; and performing inverse normalization and differencing on the second test data to obtain the second load data.

[0020] A second aspect of the present invention provides a load prediction system that integrates dual-mode decomposition and attention mechanisms, the system comprising: The first data decomposition unit is used to remove high-frequency noise from the first load signal by cascading processing of Complete Ensemble Empirical Mode Decomposition (CEEMD) and Variational Mode Decomposition (VMD), including: performing data decomposition on the first load signal through CEEMD to obtain a residual signal; inputting the residual signal into VMD to perform local frequency band decomposition on the residual signal to obtain multi-scale feature components.

[0021] The second data processing unit is used to reorganize the multi-scale feature components according to the time series to form a multi-dimensional input sample matrix; the multi-dimensional input sample matrix includes a first sample matrix and a second sample matrix.

[0022] The model training unit is used to sequentially input the first sample matrix into a bidirectional long short-term memory network (Bi-LSTM) for model training and obtain the model training results.

[0023] The prediction unit is used to predict the second sample matrix using the trained Bi-LSTM model to obtain the second load data.

[0024] According to an embodiment of the present invention, the first data decomposition subunit includes: a noise injection module, used to inject different Gaussian noise sequences into a first load signal to obtain a noisy signal set; and a first decomposition module, used to perform empirical mode decomposition on each group of signals in the noisy signal set and calculate the mean of each intrinsic mode function, and peel off high-frequency components layer by layer to obtain a residual signal.

[0025] According to one embodiment of the present invention, the cascading of CEEMD and VMD adopts an adaptive threshold determination mechanism, using the rate of change of the spectral density of the residual signal as the termination condition for the decomposition depth, and dynamically determining the number of decomposition layers.

[0026] According to an embodiment of the present invention, the model training unit includes: a first calculation module, used to calculate the forward hidden state vector and the backward hidden state vector respectively through a dual-channel gating unit, and to calculate the attention score by combining the current context information through a gating mechanism.

[0027] The second calculation module is used to normalize the attention score to obtain the attention weight at each time step; and to obtain the context vector by weighted summation.

[0028] The setting module is used to set the iteration number threshold and the time step prediction result error function, wherein the time step prediction result error function is expressed as the average value of the prediction result errors of all time steps within the time period; and to set the average threshold of the time step prediction result error.

[0029] The third calculation module is used to generate a first test result based on the context vector and compare it with the first sample matrix. Based on the comparison result, it determines whether training has ended, including: if the difference between the first test result and the data in the first sample matrix is ​​less than the average threshold of the prediction result error of all time steps within a preset time period, and is greater than the iteration number threshold, training ends; if the difference between the first test result and the data in the first sample matrix is ​​not less than the average threshold of the prediction result error of all time steps within a preset time period, or is not greater than the iteration number threshold, it returns to the first calculation module to continue iterative training.

[0030] A third aspect of the present invention provides an intelligent device, including a transmitter, a receiver, a memory, and a processor; the memory is used to store computer instructions; the processor is used to execute the computer instructions stored in the memory to implement the above-mentioned load prediction method that integrates dual-mode decomposition and attention mechanisms.

[0031] A fourth aspect of the present invention provides a storage medium, comprising: a readable storage medium and computer instructions, wherein the computer instructions are stored in the readable storage medium; the computer instructions are used to implement the above-described load prediction method that integrates dual-mode decomposition and attention mechanisms.

[0032] Compared to existing technologies, this invention offers the following advantages: First, by employing CEEMDAN's adaptive ensemble noise-assisted decomposition to perform global coarse-grained screening of the load sequence, followed by VMD for multi-component fine decomposition in the frequency domain, it not only suppresses pseudo-modes and mode aliasing but also extracts multi-scale information such as trends, periods, transients, and random disturbances within the same process. These complementary features, input hierarchically in subsequent modeling stages, significantly enrich the representational dimensions of training samples, thereby helping the model identify complex load patterns. Second, by utilizing forward LSTM to mine the impact of long-term historical states on the current moment, and backward LSTM to capture the constraints of future evolution on previous states, and then adaptively adjusting the weights of bidirectional representations in the time dimension through a dynamic attention mechanism, it can highlight attention to abnormal segments such as peaks and valleys, and reduce prediction bias caused by temporal non-stationarity, while maintaining continuous characterization of normal segments. Furthermore, the concatenated decomposition and bidirectional attention framework achieves higher prediction accuracy by improving feature utilization and temporal characterization precision; it significantly outperforms mainstream models in six commonly used error metrics, including MAE, RMSE, MAPE, SMAPE, R², and Theil index, demonstrating stronger generalization ability and robustness in various operating scenarios. Attached Figure Description

[0033] Figure 1 This is a flowchart of the load prediction method that integrates dual-mode decomposition and attention mechanism disclosed in an embodiment of the present invention; Figure 2 This is an example diagram of initial load data disclosed in an embodiment of the present invention; Figure 3(a) is an example of data obtained by secondary decomposition of the initial load data IMF1 after the first decomposition component of the present invention disclosed in the embodiment of the present invention; Figure 3(b) is an example of data obtained by secondary decomposition of the initial load data IMF2 after the first decomposition component, as disclosed in the embodiment of the present invention. Figure 3(c) is an example of data obtained by secondary decomposition of the initial load data IMF3 after the first decomposition component, as disclosed in the embodiment of the present invention. Figure 3(d) is an example of data obtained by secondary decomposition of the initial load data IMF4 after the first decomposition component, as disclosed in the embodiment of the present invention. Figure 4 This is a predicted load fitting curve diagram disclosed in an embodiment of the present invention; Figure 5 This is a curve showing the absolute error of load forecasting as disclosed in an embodiment of the present invention. Figure 6 This is a block diagram of a load prediction system that integrates dual-mode decomposition and attention mechanisms, as disclosed in an embodiment of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings. 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. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application.

[0035] The first aspect of this invention provides a load prediction method that integrates bimodal decomposition and attention mechanisms, such as... Figure 1 As shown.

[0036] First, the raw load data undergoes dual-mode decomposition and adaptive noise complete ensemble empirical mode decomposition. The raw data is first cleaned, and linear interpolation is used to fill in any missing values. CEEMDAN (Adaptive Noise Complete Ensemble Empirical Mode Decomposition) suppresses mode aliasing through adaptive noise injection and ensemble averaging. Its core formula is: Equation (1) This is the original load sequence. This represents the noise intensity (usually set to 0.2 times the signal standard deviation). It is Gaussian white noise. Add the noise a certain number of times (usually 100 times); Indicates the first EMD decomposition results of the first-order IMF.

[0037] CEEMDAN first performs noise injection, adding N different noise sequences to the original signal x(t), generating a noisy signal set {x+ If wi}i=1N, then perform layered decomposition, perform EMD decomposition on each group of noisy signals, calculate the IMF mean layer by layer, and finally use the residual of the previous IMF as the input for the next decomposition until the residual is a monotonic function.

[0038] In the experiment, CEEMDAN decomposed the original load data into Kc groups of IMFs (Intrinsic Mode Functions): CEEMDAN-IMF1, CEEMDAN-IMF2, CEEMDAN-IMF3, ... CEEMDAN-IMFKc.

[0039] Variational mode decomposition (VMD) is then performed, and VMD achieves precise signal separation through constrained variational optimization. Its mathematical model is as follows: Equation (2) in, For the first One modal function, For the center frequency, The preset number of modes.

[0040] VMD is first initialized, setting the center frequencies of each mode. With Lagrange multipliers, and then alternately updated using the ADMM algorithm. and until the following constraints are met: Equation (3) VMD performs a secondary decomposition on each CEEMDAN-IMF component to obtain the corresponding Kv component amount.

[0041] The synergy between CEEMDAN global coarse screening and VMD local fine segmentation can resolve the spectral overlap problem inherent in single decomposition. Through secondary decomposition, the original load data can be further refined... Figure 2 Decompose it to get × The component quantities are shown in Figures 3(a), 3(b), 3(c), and 3(d). The original signal is decomposed using CEEMDAN to obtain multiple IMF components, and each IMF component is further decomposed using VMD to achieve fine separation of signals in different frequency bands. IMF1 mainly contains high-frequency impulse information, while IMF4 reflects low-frequency trend information. The layer-by-layer decomposition results, from the most detailed IMF1 to the most trending IMF4, show that each layer is further subdivided into different frequency sub-signals by VMD.

[0042] The data is differencing and subjected to supervised transformation. First-order differencing is then performed on the decomposed datasets to reduce trend and dimensionality, making the load data more stable and easier for the model to learn. The data is then normalized to between 0 and 1 to accelerate model training convergence, while the scaler used during normalization is saved for inverse scaling of the output. A supervised transformation is then performed, converting each column of load data into double-column misaligned supervised data, facilitating the model's learning of the time-series mapping of load data. Finally, the dataset is divided into non-overlapping training and test sets based on time.

[0043] According to another embodiment of the present invention, after signal decomposition, the multi-scale feature components obtained by complete ensemble empirical mode decomposition (CEEMDAN) and variational mode decomposition (VMD) are reorganized and preprocessed to form a multi-dimensional input sample matrix that the model can learn. The step is characterized by including the following sub-steps: The load feature components obtained by CEEMDAN–VMD decomposition at each scale are aligned and paired according to a unified time index. The multiple components corresponding to the same time are combined into a set of feature vectors to construct a multi-dimensional sample matrix with time step as the row and component type as the column, which is used for subsequent learning.

[0044] For each time series component in the multidimensional sample matrix, the differencing order is automatically determined based on the autocorrelation coefficient and stationarity statistics (such as the ADF test result and the rate of change of variance). Higher-order differencing is performed when a significant trend is detected; the original sequence is kept unchanged when the sequence already meets the stationarity condition. This achieves adaptive stationarization for different frequency components, preventing information loss caused by over-smoothing.

[0045] A time-sliding window method is used to calculate the maximum and minimum values ​​within a local interval. Dynamic normalization is then performed on the differencing samples, allowing the normalization interval to adaptively adjust over time. When load fluctuations are severe, the normalization interval is automatically expanded to prevent gradient saturation. Simultaneously, the scaling function of the normalization parameters over time is saved to enable inverse normalization restoration for the corresponding time period during the prediction phase, ensuring that the output data is consistent with the actual load dimensions.

[0046] When constructing time-series supervised samples, the input-output samples are weighted according to the energy density, information entropy, or contribution weight allocation coefficient of each scale component. The multidimensional input features at time step t are mapped to the target load value at time step (t+1) to form a double-column misaligned supervised data matrix. This enhances the model's ability to learn high-contribution components and key load change patterns.

[0047] The supervised samples are divided according to their chronological order to generate training and test sets. At the same time, a time-related evaluation function (such as similarity entropy or correlation coefficient threshold) is used to constrain the consistency of the distribution of the training and test sets, ensuring that the training and validation samples remain statistically comparable and preventing time leakage.

[0048] Through the aforementioned multi-scale sample matrix adaptive preprocessing steps, trend terms and non-stationary fluctuations can be dynamically suppressed, improving the learnability of load features. Time-adaptive adjustment of feature scales enhances the model's robustness to periods of drastic fluctuation. Weighted supervision construction improves the expression strength of key components. Ensuring the consistency of training and testing data distribution improves prediction accuracy and generalization ability.

[0049] The model is then trained using the training set. The model includes a Bidirectional Long Short-Term Memory (BiLSTM) network, which simultaneously extracts temporal features through forward and backward dual-channel LSTMs, overcoming the limitations of traditional LSTMs' unidirectional information transmission. Its advantage lies in the ability of bidirectional LSTMs to simultaneously capture both historical dependencies and future trends in workload data (such as sudden load spikes during holidays). An attention mechanism is then incorporated to enhance the model's ability to learn temporal features. Its output is: (Formula 4) Input features (decomposed IMF / VMD components) at time t. and These represent the forward / backward hidden states, respectively. Let be the weight matrix, and ⊕ denotes vector concatenation.

[0050] After setting the number of iterations (Epoch), within each Epoch, calculate the error parameter MSE of the prediction results of all time steps according to (Formula 13), and calculate the average value of the MSE of the prediction results of all time steps in that period as the loss function (Loss). When the Loss decreases to a stable level and the number of iterations is greater than or equal to the number of Epochs, the training is considered complete. If the two conditions cannot be met at the same time, continue training and adjust the internal parameters of the model with the goal of minimizing the Loss until the set conditions are met. After the training is completed, save the model for prediction experiments.

[0051] Specific weights wA are assigned by calculating the correlation between inputs obtained asynchronously. This parameter quantifies the correlation between inputs at different time steps. In the attention layer, the output h of the previous layer is defined as the input xA of this layer, with weights WA, n as the time step, bA as the bias, and z as the attention component. The formula for calculating the attention component is: Equation (5) Subsequently, the attention components are normalized using the Softmax function to obtain the component wi at the i-th time step in the attention weight wA, that is: Equation (6) Finally, the context vector c is calculated one by one, that is: Equation (7) in yes The The time step component has the following trainable parameters. , .

[0052] In the test set, prediction experiments are conducted using the trained model. The model will output the prediction results for each time step in the test set in sequence. However, these results are predicted values ​​after differencing and normalization. Therefore, it is necessary to use the saved scaler to inversely normalize the predicted values ​​and then perform inverse differencing to obtain the final load prediction values.

[0053] The experiment used six metrics to measure the model's prediction accuracy. Mean Absolute Error (MAE), also known as L1 norm loss, is the average of the absolute values ​​of the prediction errors, representing the average difference between the predicted and observed values. A smaller MAE value indicates more accurate predictions. Its calculation formula is: Equation (8) Mean Absolute Percentage Error (MAPE) is the average percentage of prediction error, reflecting the average deviation of predicted values ​​from actual values. A smaller MAPE value indicates better model performance. Its calculation formula is: Equation (9) The root mean square error (RMSE) is the square root of the average prediction error, emphasizing sensitivity to large errors. A smaller RMSE value indicates more accurate model predictions. Its formula is: Equation (10) Normalized Root Mean Square Error (nRMSE) is calculated by dividing the RMSE by the difference between the maximum and minimum true values. It normalizes the error to a relative range; a smaller nRMSE value indicates more accurate model predictions. Its calculation method is as follows: Equation (11) The R-squared parameter (R²) is a quantification of how well a model explains the population variance; it is also known as the coefficient of determination or goodness of fit, and ranges from [0,1]. An R² value closer to 1 is better; 1 indicates a perfect fit, 0 indicates the model cannot explain the target variable, and a negative value indicates a poor fit. Its calculation formula is: Equation (12) MSE, also known as L2 norm loss, is the summation and averaging of the squared prediction errors. In load forecasting model training, we chose MSE as the objective function. The formula for calculating MSE is: Equation (13) In the formula: These are the model's predicted values; The actual value; It is the average of real historical data; To predict the length of the sequence.

[0054] The experiment used load data from the ISO public load dataset provided by an independent system operator in a certain region for verification testing. The dataset contains the sum of the 2016 loads of eight regions in the western and central parts of the region. The total load for the first 11 months was selected as the training set, and the total load for the last 12 months was selected as the test set. The experimental parameters are shown in Table 1.

[0055] Table 1 Experimental parameter settings

[0056] The comparison of prediction error indices corresponds to the experimental results in Table 1 above, as shown in Table 2, and Table 3.

[0057] Table 2 Comparison of Prediction Result Error Indicators 1

[0058] Table 3 Comparison of Prediction Result Error Indicators 2

[0059] As shown in Tables 2 and 3, the CV-BL-A model significantly outperforms the LSTM and RNN models across all error metrics. Specifically: in terms of MSE, CV-BL-A reduces the error by 44.3% compared to LSTM and 49.8% compared to RNN; in terms of MAE, CV-BL-A reduces the error by 14.0% compared to LSTM and 21.0% compared to RNN; in terms of RMSE, CV-BL-A reduces the error by 25.4% compared to LSTM and 29.2% compared to RNN; in terms of nMSE, CV-BL-A reduces the error by 44.4% compared to LSTM; in terms of nRMSE, CV-BL-A reduces the error by 22.6% compared to LSTM; and in terms of MAPE, CV-BL-A reduces the error by 9.5% compared to LSTM and 16.7% compared to RNN. Furthermore, the R² value of CV-BL-A is as high as 0.991, which is 0.006 and 0.008 higher than LSTM and RNN respectively, approaching the theoretical optimum of 1, indicating that the model can explain 99.1% of the variance in the data.

[0060] Figure 4 The load forecast fitting curve is shown in the image. The forecast curve fits the actual load curve very well, and it can follow the load changes well at the maximum and minimum points. Figure 5 This represents the absolute error between the predicted and actual values. It can be observed that when the load dimension is 20000MW, the absolute error of the load forecast can be kept below 750MW, which can achieve stable and effective forecasting. The stability of the forecasting performance is very important for the application scenarios of load forecasting.

[0061] As shown in Table 4, the comparison of training time costs reveals that increasing the decomposition of IMF data (CEEMDAN decomposes it into 8 groups, VMD into 3 more groups, plus 1 group of load values) does not significantly increase the training time of the CV-BL-A model per run. The proposed model achieves better prediction results with almost the same training time cost.

[0062] Table 4 Comparison of Training Duration

[0063] A second aspect of the present invention provides a load prediction system 70 that integrates dual-mode decomposition and attention mechanisms, such as... Figure 6 As shown, it includes: a first data decomposition unit 701, used to remove high-frequency noise from the first load signal through cascaded processing of complete set empirical mode decomposition and variational mode decomposition; the first data decomposition subunit is used to decompose the first load signal using the complete set empirical mode decomposition algorithm to obtain a residual signal; the second data decomposition subunit is used to input the residual signal to the variational mode decomposition module to perform local frequency band decomposition on the residual signal to obtain multi-scale feature components.

[0064] The second data processing unit 702 is used to reorganize the multi-scale feature components according to the time series to form a multi-dimensional input sample matrix, wherein the multi-dimensional input sample matrix includes a first sample matrix and a second sample matrix.

[0065] The model training unit 703 is used to input the first sample matrix into the bidirectional long short-term memory network in sequence to train the model and obtain the model training result.

[0066] The prediction unit 704 is used to take the second sample matrix as input and obtain the second load data according to the trained model.

[0067] According to an embodiment of the present invention, the first data decomposition subunit includes: a noise injection module, used to inject different Gaussian noise sequences into a first load signal to obtain a noisy signal set; and a first decomposition module, used to perform empirical mode decomposition on each group of signals in the noisy signal set and calculate the mean of each intrinsic mode function, and peel off high-frequency components layer by layer to obtain a residual signal.

[0068] According to one embodiment of the present invention, the cascade of the complete set empirical mode decomposition and variational mode decomposition adopts an adaptive threshold determination mechanism, using the rate of change of the spectral energy density of the residual signal as the termination condition for the decomposition depth, and dynamically determining the number of decomposition layers.

[0069] According to an embodiment of the present invention, the model training unit includes: a first computation module, configured to compute the forward hidden state vector and the backward hidden state vector respectively through a dual-channel gating unit, and to compute an attention score by combining the current context information through a gating mechanism. The second calculation module is used to normalize the attention score to obtain the attention weight at each time step; and to obtain the context vector by weighted summation.

[0070] The setting module is used to set the iteration number threshold and the time step prediction result error function, wherein the time step prediction result error function is expressed as the average value of the prediction result errors of all time steps within the time period; and to set the average threshold of the time step prediction result error. The third calculation module is used to generate a first test result based on the context vector and compare it with the first sample matrix; determine whether training has ended based on the comparison result. If the difference between the first test result and the data in the first sample matrix is ​​less than the average threshold of the prediction result error of all time steps within a preset time period, and is greater than the iteration number threshold, training ends; if the difference between the first test result and the data in the first sample matrix is ​​not less than the average threshold of the prediction result error of all time steps within a preset time period, or is not greater than the iteration number threshold, return to the first calculation module and continue iterative training.

[0071] A third aspect of the present invention provides an intelligent device, including a transmitter, a receiver, a memory, and a processor; the memory is used to store computer instructions; the processor is used to execute the computer instructions stored in the memory to implement the above-mentioned load prediction method that integrates dual-mode decomposition and attention mechanisms.

[0072] A fourth aspect of the present invention provides a storage medium comprising: a readable storage medium and computer instructions, wherein the computer... The instructions are stored in the readable storage medium; the computer instructions are used to implement the above-mentioned fusion dual-mode decomposition and attention mechanism. Load forecasting methods.

[0073] Compared to existing technologies, this invention offers the following advantages: First, by employing CEEMDAN's adaptive ensemble noise-assisted decomposition to perform global coarse-grained screening of the load sequence, followed by VMD for multi-component fine decomposition in the frequency domain, it not only suppresses pseudo-modes and mode aliasing but also extracts multi-scale information such as trends, periods, transients, and random disturbances within the same process. These complementary features, input hierarchically in subsequent modeling stages, significantly enrich the representational dimensions of training samples, thereby helping the model identify complex load patterns. Second, by utilizing forward LSTM to mine the impact of long-term historical states on the current moment, and backward LSTM to capture the constraints of future evolution on previous states, and then adaptively adjusting the weights of bidirectional representations in the time dimension through a dynamic attention mechanism, it can highlight attention to abnormal segments such as peaks and valleys, and reduce prediction bias caused by temporal non-stationarity, while maintaining continuous characterization of normal segments. Furthermore, the concatenated decomposition and bidirectional attention framework achieves higher prediction accuracy by improving feature utilization and temporal characterization precision; it significantly outperforms mainstream models in six commonly used error metrics, including MAE, RMSE, MAPE, SMAPE, R², and Theil index, demonstrating stronger generalization ability and robustness in various operating scenarios.

[0074] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A load prediction method integrating dual-mode decomposition and attention mechanism, characterized in that, The method is used to predict second load data from first load signal data, including: The first load signal is processed by cascading CEEMD and VMD to remove high-frequency noise, including: performing data decomposition on the first load signal by CEEMD to obtain a residual signal; inputting the residual signal into VMD to perform local frequency band decomposition on the residual signal to obtain multi-scale feature components; the cascading of CEEMD and VMD adopts an adaptive threshold determination mechanism, using the rate of change of the spectral energy density of the residual signal as the termination condition for the decomposition depth, and dynamically determining the number of decomposition layers; The multi-scale feature components are reorganized according to the time series to form a multi-dimensional input sample matrix; the multi-dimensional input sample matrix includes a first sample matrix and a second sample matrix, the first sample matrix and the second sample matrix include time-aligned trend features, periodic features and transient disturbance components, which are used to represent the multi-layer feature information of the load time series signal; The first sample matrix is ​​sequentially input into a Bi-LSTM for model training, including: extracting bidirectional hidden state vectors based on the potential correlation between historical information and future data modeled in the forward and backward channels of the Bi-LSTM; constructing a dynamic attention weight matrix based on the bidirectional hidden state vectors; calculating the importance scores of features at each time step through an adaptive weight adjustment mechanism; and performing weighted fusion in the time dimension to obtain a global temporal feature representation, followed by iterative calculation; including: Step S1: Calculate the forward hidden state vector and the backward hidden state vector using a dual-channel gating unit, and calculate the attention score by combining the gating mechanism with the current context information. Step S2: Normalize the attention score to obtain the attention weight at each time step; obtain the context vector by weighted summation. Step S3: Set the iteration number threshold and the time step prediction result error function. The time step prediction result error function is expressed as the average value of the prediction result errors of all time steps within a preset time period; set the average threshold of the time step prediction result error. Step S4: Generate a first test result based on the context vector and compare it with the first sample matrix. Determine whether training is complete based on the comparison result, including: If the difference between the first test result and the first sample matrix data is less than the average threshold of the prediction result error of all time steps within the preset time period, and greater than the iteration number threshold, the training ends. If the difference between the first test result and the first sample matrix data is not less than the average threshold of the prediction result error of all time steps within the preset time period, or is not greater than the iteration number threshold, return to S1; The trained Bi-LSTM model is used to predict the second sample matrix to obtain the second load data.

2. The method according to claim 1, characterized in that, The step of performing data decomposition on the first load signal using CEEMD to obtain a residual signal includes: injecting different Gaussian noise sequences into the first load signal to obtain a noisy signal set; performing empirical mode decomposition on each signal in the noisy signal set and calculating the mean of each intrinsic mode function, and peeling off high-frequency components layer by layer to obtain the residual signal.

3. The method according to claim 1, characterized in that, The step of using the trained Bi-LSTM model to predict the second sample matrix to obtain the second load data includes: Using the second sample matrix as input, and based on the trained model, the second test data is obtained. The second test data is inversely normalized and differentially processed to obtain the second load data.

4. A load forecasting system integrating dual-mode decomposition and attention mechanism, characterized in that, The system for predicting second load data from first load signal data includes: The first data decomposition unit is used to remove high-frequency noise from the first load signal through cascaded processing of CEEMD and VMD. This includes: performing data decomposition on the first load signal using CEEMD to obtain a residual signal; inputting the residual signal into VMD to perform local frequency band decomposition on the residual signal to obtain multi-scale feature components; the cascaded processing of CEEMD and VMD employs an adaptive threshold determination mechanism, using the rate of change of the spectral energy density of the residual signal as the termination condition for the decomposition depth, and dynamically determining the number of decomposition layers. The second data processing unit is used to reorganize the multi-scale feature components according to the time series to form a multi-dimensional input sample matrix; the multi-dimensional input sample matrix includes a first sample matrix and a second sample matrix. The model training unit is used to sequentially input the first sample matrix into the Bi-LSTM for model training and obtain the model training results. The prediction unit is used to predict the second sample matrix using the trained Bi-LSTM model to obtain the second load data; The model training unit includes: The first calculation module is used to calculate the forward hidden state vector and the backward hidden state vector respectively through a dual-channel gating unit, and to calculate the attention score by combining the gating mechanism with the current context information. The second calculation module is used to normalize the attention score to obtain the attention weight at each time step; and to obtain the context vector by weighted summation. The setting module is used to set the iteration number threshold and the time step prediction result error function, wherein the time step prediction result error function is expressed as the average value of the prediction result errors of all time steps within a preset time period; and to set the average threshold of the time step prediction result error. The third calculation module is used to generate a first test result based on the context vector and compare it with the first sample matrix. Based on the comparison result, it determines whether training has ended, including: if the difference between the first test result and the data in the first sample matrix is ​​less than the average threshold of the prediction result error of all time steps within a preset time period, and is greater than the iteration number threshold, training ends; if the difference between the first test result and the data in the first sample matrix is ​​not less than the average threshold of the prediction result error of all time steps within a preset time period, or is not greater than the iteration number threshold, it returns to the first calculation module to continue iterative training.

5. The system according to claim 4, characterized in that, The first data decomposition unit includes: a noise injection module, used to inject different Gaussian noise sequences into the first load signal to obtain a noisy signal set; and a first decomposition module, used to perform empirical mode decomposition on each group of signals in the noisy signal set and calculate the mean of each order intrinsic mode function, and peel off high-frequency components layer by layer to obtain residual signals.

Images