Copula entropy and VMD-BiTCN-LSTM fused electrical load prediction method, system and device, and storage medium

Abstract

The invention discloses a Copula entropy and VMD-BiTCN-LSTM fused electrical load prediction method, a Copula entropy and VMD-BiTCN-LSTM fused electrical load prediction system, equipment and a storage medium. The method comprises the following steps: firstly, acquiring a data set containing a historical power load sequence and a plurality of original feature sequences, and deeply mining and screening out a target feature sequence having strong nonlinear correlation with the load sequence from original features by using a feature screening method based on Copula entropy; thirdly, decomposing a non-stationary historical load sequence into a plurality of intrinsic mode function subsequences containing long-term trends, periodic fluctuations and random components by adopting a variational mode decomposition (VMD) algorithm, and fusing the intrinsic mode function subsequences with the screened target feature sequence to form a new data set; and finally, inputting the fused data set into a composite prediction model which is sequentially connected in series with a bidirectional time sequence convolutional network BiTCN, a time sequence attention mechanism TPA and a long short-term memory network LSTM, and generating a high-precision short-term power load prediction result through parallel extraction of dynamic features by the BiTCN, focusing of key period information by the TPA and capturing of a long-term dependency relationship by the LSTM.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system automation, and in particular to a short-term power load prediction method, system, device and computer readable storage medium which fuses advanced feature screening, signal decomposition and deep learning model. BACKGROUND

[0002] As a core link for ensuring the safe and stable operation and economic and efficient dispatch of the power grid system, the accuracy of the prediction result of short-term power load prediction has a crucial influence on formulating a reasonable power generation plan, optimizing energy dispatch, reducing operation cost and maintaining power grid safety margin. Traditional short-term load prediction techniques mostly rely on time series analysis, regression analysis and other classical statistical models. Although these methods have relatively simple model construction, their inherent linear assumption makes them appear inadequate when facing modern power systems. Current power load data is deeply influenced by the complex interweaving of multi-dimensional factors such as meteorological mutations, holiday effects, social and economic activities and various random events, showing strong nonlinearity, high volatility and time-varying dynamic characteristics, which makes it difficult for traditional statistical models to effectively capture the inherent law, and there are significant bottlenecks in prediction accuracy and generalization ability, which cannot meet the urgent needs of modern power grid fine and intelligent management.

[0003] To cope with the above challenges, machine learning and deep learning technologies represented by artificial neural networks, support vector machines and especially long short-term memory networks (LSTM) are introduced into the field of power load prediction. These technologies have made great progress in prediction accuracy due to their excellent non-linear data fitting capability. However, as research deepens, these advanced methods have also exposed new and deeper technical problems.

[0004] Firstly, at the feature engineering level, the number of factors affecting power load is large and high-dimensional, and they are not simply linearly superimposed, but there are complex nonlinear coupling and high-order dependence relationships. Existing feature selection methods, such as Pearson correlation coefficient analysis, mainly measure the linear correlation between variables, and when dealing with such complex relationships, they often cannot accurately identify key nonlinear influencing factors, and may even introduce redundant features that interfere with model training, thereby limiting the upper limit of model performance from the data source.

[0005] Secondly, at the level of signal processing, the original power load time series is essentially a non-stationary signal, which mixes long-term trends that change slowly, fluctuation components with different periods, and irregular random noise. If this complex mixed signal is directly input into a single prediction model, it will undoubtedly greatly increase the learning burden of the model, making it difficult to converge to the optimal state, especially when predicting the mutation point of the load. Although the industry has tried to use wavelet transform or empirical mode decomposition to pre-decompose the signal, these methods have inherent defects such as the need for manual selection of the optimal basis function, or the easy production of mode aliasing and destruction of component physical meaning. As a more advanced signal decomposition technique, VMD alleviates the above problems to some extent, but how to effectively integrate it with the downstream prediction model is still a direction to be explored.

[0006] Finally, at the level of prediction model construction, the single deep learning model structure also has its ability boundary. For example, LSTM network is good at capturing long-term dependencies of time series, but it is not good at efficiently extracting local fine-grained dynamic features; while the convolutional model represented by time convolution network (TCN) can efficiently capture local features through parallel computing, but it is insufficient in establishing global long-term dependencies. Therefore, simply combining VMD decomposition with a single LSTM or TCN model is still difficult to comprehensively and cooperatively handle the respective characteristics of the decomposed components of different scales, and cannot achieve a complete characterization of the dynamic changes of power load.

[0007] In summary, constructing a comprehensive prediction framework that can accurately reduce features from the root, finely process the non-stationarity of the signal, and integrate the advantages of multiple models to achieve all-around feature learning is a technical problem that needs to be overcome in the current power load prediction field. SUMMARY

[0008] The present application aims to solve the problem of the single dimension of feature selection, poor handling of non-stationarity of load sequence, limited ability of single prediction model, and insufficient overall prediction accuracy and stability in the prior art short-term power load prediction method.

[0009] To solve the above problems, in a first aspect, the present application provides a power load prediction method combining Copula entropy and VMD-BiTCN-LSTM, comprising the following steps: S1: Obtain a data set containing a historical power load sequence and a plurality of original feature sequences, and use a feature selection method based on Copula entropy to select a target feature sequence with strong correlation with the historical power load sequence from the plurality of original feature sequences; S2: adopting a variational mode decomposition (VMD) algorithm to decompose the historical power load sequence into a plurality of intrinsic mode function (IMF) sub-sequences, and fusing the plurality of IMF sub-sequences with the target feature sequence screened in step S1 to obtain a fusion data set; S3: adopting a bidirectional time convolution network (BiTCN) to perform bidirectional feature extraction on the fusion data set to obtain dynamic features; S4: adopting a time series attention mechanism (TPA) to weight the dynamic features to enhance load cycle characteristics to obtain weighted features; S5: inputting the weighted features into a long short-term memory (LSTM) network for time series learning and generating a short-term power load prediction result.

[0010] Preferably, the feature screening method based on Copula entropy in step S1 specifically comprises: S11: for any original feature sequence and the historical power load sequence, calculating a Copula density function ; S12: calculating a Copula entropy corresponding to the Copula density function; S13: according to the principle that the smaller the Copula entropy is, the stronger the correlation is, selecting an original feature sequence with a Copula entropy less than a preset threshold as the target feature sequence.

[0011] Preferably, after step S2 and before step S3, the method further comprises: S21: pre-processing the fusion data set, wherein the pre-processing comprises linear interpolation filling of missing values, correction of abnormal values, and data normalization.

[0012] Preferably, after step S5, the method further comprises: mapping the output of the long short-term memory (LSTM) network through a fully connected layer to generate the short-term power load prediction result.

[0013] Preferably, in step S3, the bidirectional time convolution network (BiTCN) divides and parallelly processes the fusion data set in a sliding window manner.

[0014] Preferably, in step S5, the internal calculation process of the long short-term memory (LSTM) network comprises: wherein , are the hidden layer outputs at the previous time and the current time, respectively; is the input at the current time; , , are weights of the forget gate, update gate, and output gate, respectively; is a candidate value vector; , are cell states of the previous time and the current time, respectively; and is an activation function.

[0015] In a second aspect, the present application provides an electric load prediction system fusing Copula entropy and VMD-BiTCN-LSTM, which is configured to perform the method of the first aspect, comprising: a feature screening module configured to perform steps S11-S13 of claim 2 to screen a target feature sequence from a dataset containing a historical electric load sequence and a plurality of original feature sequences; a sequence processing module connected to the output end of the feature screening module and configured to decompose the historical electric load sequence into a plurality of intrinsic mode function (IMF) sub-sequences using a variational mode decomposition (VMD) algorithm, and fuse the IMF sub-sequences with the target feature sequence to generate a fusion dataset; a prediction module having an input end connected to the output end of the sequence processing module, configured to receive the fusion dataset and generate a short-term electric load prediction result; wherein the prediction module internally comprises, in sequence: a bidirectional time convolutional network (BiTCN) submodule for extracting dynamic features; a time-based attention mechanism (TPA) submodule having an input end connected to the output end of the BiTCN submodule, configured to weight the dynamic features; and a long short-term memory (LSTM) submodule having an input end connected to the output end of the TPA submodule, configured to generate a final prediction result.

[0016] Preferably, the prediction module further comprises a fully connected layer submodule having an input end connected to the output end of the LSTM submodule, configured to map the output of the LSTM submodule.

[0017] In a third aspect, the present application provides a computer device comprising a processor and a memory, wherein the memory stores a computer program, and the processor, when executing the computer program, implements the method of the first aspect.

[0018] In a fourth aspect, the present application further provides a computer-readable storage medium storing a computer program, wherein the program, when executed by a processor, implements the method of the first aspect.

[0019] The beneficial effects of the present application are: Firstly at the data input end, the present application adopts Copula entropy for feature screening, and the core advantage thereof lies in the ability to penetrate the data appearance and accurately measure the deep nonlinear dependence relationship between variables, thereby ensuring that the model receives the core driving factors that are truly strongly related to load changes, which fundamentally avoids the interference of redundant data on the prediction accuracy and lays a solid data foundation for subsequent accurate prediction.

[0020] On the basis of high-quality input, in order to cope with the inherent non-stationarity and complexity of the load sequence, the present application further adopts the VMD algorithm to finely decompose the original mixed signal into long-term trend, periodic fluctuation and random component with clear physical meaning, which greatly reduces the learning difficulty of a single model, so that the subsequent model can learn in a special and efficient manner according to the characteristics of different components, and significantly improves the adaptability and capturing ability of the model to the dynamic changes of the load.

[0021] Finally, the prediction core of the present application is a specially constructed composite model with deep synergy, wherein the BiTCN is responsible for efficiently extracting the local dynamic details of the sequence from the front and back directions in parallel; the TPA that follows is like an intelligent "focus" that dynamically assigns weights to the features at different time points, so that the model can focus on processing key periodic information such as load peaks and valleys; and these deeply refined and focused features are finally sent to the LSTM, which uses a powerful gating memory mechanism to capture long-term dependencies throughout.

[0022] It is through the seamless integration and synergistic gain of the three stages of ensuring the "quality" of the input by Copula entropy, ensuring the "purity" of the sequence by VMD decomposition, and ensuring the "precision" of processing by the BiTCN-TPA-LSTM composite model that the overall framework proposed by the present application exhibits superior performance far beyond traditional models and simple combination models in key performance indicators such as MAPE and RMSE, and achieves significant improvement in the accuracy and stability of the prediction results. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a general flow framework diagram of a short-term power load prediction method provided by an embodiment of the present application.

[0024] Figure 2 is a column chart of feature factor and load dependence strength ranking based on Copula entropy in an embodiment of the present application.

[0025] Figure 3This is a diagram showing the result of the original load sequence being decomposed by the VMD algorithm in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the internal structure of the BiTCN module in an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram of the internal structure of the LSTM unit in an embodiment of the present invention.

[0028] Figure 6 This is a comparison chart of the prediction results of the model proposed in this embodiment of the invention with other comparative models. Detailed Implementation

[0029] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0030] Example 1: This embodiment details a power load forecasting method that integrates Copula entropy and VMD-BiTCN-LSTM. This method aims to significantly improve the accuracy and robustness of short-term power load forecasting through an end-to-end process that integrates advanced feature selection, refined signal decomposition, and a deep collaborative forecasting model. (Refer to...) Figure 1 The overall process framework of this method clearly shows the key steps from raw data input to final prediction output. Each step in this method will be explained in detail below.

[0031] First, this method performs data acquisition and feature selection step S1. Its core lies in accurately identifying the key driving features with the strongest nonlinear correlation to power load changes from a massive pool of potential influencing factors. In actual power systems, load fluctuations are influenced by numerous external factors. These factors are not only numerous but also have complex nonlinear coupling relationships. Simply using all factors as model input not only causes the "curse of dimensionality," increasing the computational burden and overfitting risk, but also potentially introduces irrelevant or weakly correlated noise features that severely interfere with the model's learning process, thus limiting the final prediction accuracy. Therefore, efficient and accurate feature selection before building the prediction model is a crucial first step in improving model performance.

[0032] The original data set of the embodiment is derived from real power grid records of a large city in North China, specifically including historical power load sequences from January 1, 2020 to December 31, 2021, a total of two years, and a plurality of original weather feature sequences strictly synchronized in time with the power load sequences. Optionally, the weather feature sequences include temperature, relative humidity, wind speed, apparent temperature, sunshine duration, rainfall, and daily maximum temperature, minimum temperature, average temperature, and average wind speed calculated from the basic data. The sampling frequency of all data is 15 minutes, thereby forming a high-density time series data set.

[0033] To solve the problem that the traditional feature selection method such as Pearson correlation coefficient can only measure linear relationship and cannot effectively handle the above-mentioned complex nonlinear dependence, the application introduces a feature screening method based on Copula entropy. The essence of Copula theory is that it can decompose a joint distribution function into the marginal distribution functions of each variable and a Copula function describing the dependence structure between variables, so that the study of correlation can be independent of the specific form of the marginal distribution, which makes it have great advantages in capturing complex dependence relationships such as nonlinear, asymmetric and tail correlation. Copula entropy, as a combination of information theory and Copula theory, is an equivalent and more easily non-parametrically estimated form of mutual information, which can more robustly and accurately measure the correlation strength between variables. The specific implementation process of the feature screening method, i.e. steps S11 to S13, is designed as a rigorous quantitative evaluation process. In step S11, for each candidate original weather feature sequence in the data set and the historical power load sequence as the prediction target , the system will calculate the Copula density function between them. The density function is the second-order partial derivative of the Copula function , which is defined by the following formula: where and are the marginal distribution functions of the candidate feature and , usually and , which map the original data to the interval.

[0034] Since the probability distribution of data in the real world is usually unknown, the embodiment preferably uses a non-parametric estimation method, such as a k-nearest neighbor estimator, where the parameter k can be set empirically, for example, optionally, set k equal to 10, to robustly estimate the Copula density. Next, in step S12, based on the Copula density function obtained in the previous step, the method will calculate the Copula entropy The calculation is done by the following integral formula: ; here, is the log of the Copula density, which measures the "amount of information" contained in the dependence structure.

[0035] In information theory, the Copula entropy and the mutual information value are inverses of each other, i.e. , which means that the absolute value of the Copula entropy directly reflects how much information is contained in each other between the two variables. Therefore, in the last step S13, the method will make a decision according to the basic principle that "the smaller the Copula entropy, the stronger the correlation". By calculating and sorting the Copula entropy between all candidate features and the load sequence, a preset threshold can be selected to filter out the original feature sequence whose Copula entropy is less than the threshold; or more directly, as done in this embodiment, the top features with the smallest Copula entropy values are selected as the final target feature sequence.

[0036] Referring to Figure 2 , this figure intuitively shows the dependence strength between each meteorological feature and the power load in the form of a column chart, i.e. the sorting result of the Copula entropy. From Figure 2 , it can be clearly seen that the dependence strength of precipitation, sunshine and other factors is relatively low, while the dependence strength of wind speed, apparent temperature, humidity, temperature and daily maximum temperature is significantly higher, i.e. their Copula entropy values are the smallest. Therefore, this embodiment selects these five features to constitute the final target feature sequence, which is used for subsequent model construction. This process not only realizes efficient dimensionality reduction of input data, but more importantly, it ensures that the features input into the model are "golden features" that have the strongest and most essential nonlinear association with the prediction target, laying a solid foundation for the success of the entire prediction task.

[0037] After obtaining the core features with strong nonlinear correlation with the load, the method enters step S2, i.e. using the VMD algorithm to finely decompose the historical power load sequence itself, and then fusing the decomposed components with the target feature sequence screened out in the previous step. The reason for performing this step is that the original power load time sequence is usually a typical non-stationary signal, which internally mixes a variety of complex components of different time scales, for example, a slowly changing long-term trend determined by seasonal changes and economic development, a periodic fluctuation with daily or weekly cycles determined by people's regular work and rest activities, and irregular high-frequency random noise caused by various accidental events. If such a highly complex mixed signal is directly input into a single prediction model, the model will be difficult to effectively distinguish and learn these different characteristic components, resulting in low learning efficiency, poor model generalization ability, and especially poor prediction effect when the load mutates.

[0038] VMD is an advanced, non-recursive signal decomposition technique, which aims to decompose a real-valued input signal into a series of discrete, sparse intrinsic mode functions (IMF) with specific characteristics . Each IMF component is assumed to be an amplitude-modulated frequency-modulated signal with a limited bandwidth, with most of its frequency content compactly around a center frequency . The first step of the VMD algorithm is to construct a variational problem, first obtaining the analytic signal of each subsequence by Hilbert transform to obtain the one-sided spectrum: where is the unit impulse signal. Then, by multiplying each modal spectrum by an exponential term , the spectrum of each mode is modulated to the corresponding base frequency band. The core of the VMD algorithm is to construct and solve the following constrained variational model: ; The model aims to find a set of modal components and their corresponding center frequencies , so that the sum of the estimated bandwidths of all modal components is minimized, while ensuring that the sum of all modal components can accurately reconstruct the original signal , where is the total number of decompositions.

[0039] In order to solve this constrained variational problem, the VMD algorithm introduces a quadratic penalty term and a Lagrange multiplier , converting the problem into an unconstrained optimization problem of augmented Lagrangian expression: ; Subsequently, the alternating direction multiplier method is used to update , and alternately in the Fourier domain until a convergence condition is met.

[0040] wherein , , are the Fourier transforms of , , respectively; is the iteration number.

[0041] The iterative update formula is as follows: ; ; ; wherein is a noise tolerance parameter.

[0042] Referring to Figure 3 , the results of the original power load sequence after VMD decomposition are shown. Optionally, the number of decomposed modes in this embodiment is set to 3. As can be seen from the figure, the first intrinsic mode function IMF0 obtained by decomposition has the lowest frequency and a flat shape, which clearly depicts the overall long-term trend of the load data. The second intrinsic mode function IMF1 presents significant, high-frequency periodic fluctuations, which represent the periodic components in the load caused by factors such as day and night alternation, weekday and weekend differences, etc. The third intrinsic mode function IMF2 usually contains higher frequency random noise and glitch components. Through such decomposition, a complex prediction task is effectively decomposed into prediction tasks for trend, periodicity and noise, which are relatively simple and have single characteristics. After completing the decomposition, the method precisely aligns the obtained IMF sub-sequences with the five target feature sequences selected in step S1 on the time stamp, and then merges them into a new, more dimensionally rich and structurally clear fusion dataset, fully preparing for the input of the subsequent deep learning model.

[0043] Before formally entering the prediction model, the method also performs an important data preprocessing step, step S21. The purpose of this step is to eliminate defects in the data and the dimension difference between different features, and to ensure that the data input into the model is clean and standardized. The specific operations include: First, the missing values in the fusion dataset are filled using the linear interpolation method, which is a simple and effective method that can maintain the local trend of the sequence.

[0044] Secondly, for outliers in the dataset, such as spikes caused by data collection errors or extremely rare events, horizontal or vertical processing is used to correct them to avoid their disproportionate impact on model training. Finally, and most importantly, all data is normalized. The preferred embodiment uses the Min-Max normalization method to scale the numerical value of each feature to the interval [0, 1] by the following formula: ; wherein, is the normalized data, is the original data, and are the maximum and minimum values of the feature in the entire dataset, respectively. This step ensures that all features are on the same order of magnitude, avoiding the problem of some features dominating the gradient descent direction during model training, which helps to accelerate model convergence and improve final performance.

[0045] After all the preprocessing steps described above, the method enters the core prediction phase, steps S3, S4, and S5. These three steps together form a deep collaborative, progressive prediction model, with the internal structure sequentially linked by BiTCN, TPA, and LSTM.

[0046] First, in step S3, the fused dataset is input into the BiTCN module for preliminary feature extraction. BiTCN is a further development based on TCN. TCN itself solves the problems of slow training speed, gradient vanishing, and limited receptive field of traditional RNN and CNN by combining causal convolution, dilated convolution, and residual connection. Dilated convolution allows the convolution kernel to jump sample on the input, thereby expanding the receptive field exponentially without increasing the number of parameters and computational complexity, allowing it to capture dependencies over longer time ranges. BiTCN, on the other hand, constructs two parallel TCN channels, one processing the forward time series and the other processing the reverse time series, and then fuses the outputs of the two channels. This has the advantage that for any time point in the sequence, the model can consider not only its past information (forward TCN) but also its future information (reverse TCN), allowing it to more fully and deeply understand the context information at that time point and extract more robust dynamic features.

[0047] In this embodiment, the BiTCN module preferably uses a sliding window approach to process the input long sequence.

[0048] Specifically, for a long sequence containing each feature time series length is the load dataset, where the th feature sequence is defined as .

[0049] When the time point is the current time point, a sliding window of size can be defined, and the historical data of the feature in the window is represented as . By setting the sliding step, the window moves dynamically along the time axis, and each sliding extracts the observation values of all features in the window to form a local feature map as the input sample of the model. The BiTCN module is to extract the parallel bidirectional dynamic characteristics of the divided data set.

[0050] Referring to Figure 4 , the figure shows a basic unit structure of the TCN module, which is usually stacked by a layer of cavity convolution, a layer of weight normalization or batch normalization BN, a ReLU activation function and a layer of Dropout, and is connected through a residual connection to prevent network degradation. The BiTCN module is stacked in two directions by such basic units.

[0051] After the BiTCN efficiently extracts the deep dynamic characteristics of the sequence, these characteristics are passed to the time series attention mechanism TPA module in step S4 for weighted processing. The motivation of introducing TPA is that for a time series prediction task, not all information of the time points in the sequence has equal importance. For example, when predicting the power load at 10 o'clock the next morning, the reference value of those time points related to the load peak or trough around 10 o'clock in the past days is obviously much higher than that of other unremarkable time points.

[0052] Specifically, after the BiTCN extracts the features, a hidden state matrix is obtained. For a given prediction target (e.g. the continuous load value of the next day), TPA calculates the relevance score between the hidden state output by BiTCN at each time step and the prediction target through a scoring function .

[0053] Then, these scores are normalized by a softmax function to obtain an attention weight distribution , and the calculation formula is: ; the weight distribution clearly indicates which time points in the past the model should pay more "attention" to in the current prediction task. Finally, TPA sums up the attention weights with the corresponding hidden states to obtain an attention-focused, context-aware feature vector , whose calculation formula is Through this step, the model can automatically enhance the expression of periodic features and key time point features that have a decisive influence on the prediction result, while suppressing irrelevant information interference, so that the information input to the next layer is of higher quality and more targeted.

[0054] Finally, in step S5, this carefully refined and weighted feature vector by TPA is input into the long short-term memory network (LSTM) module for the final time series learning and prediction result generation. LSTM is a variant of recurrent neural network (RNN), and its design is intended to solve the gradient vanishing and gradient explosion problems commonly encountered by standard RNN when processing long sequences, so that it can truly learn long-term dependencies in sequences.

[0055] Referring to Figure 5 , the core of LSTM lies in its cell structure and the special mechanism called "gate" inside. A standard LSTM cell contains three key gates: forget gate , update gate and output gate . The forget gate is responsible for deciding which information should be forgotten or retained from the cell state based on the current input and the hidden state at the previous time.

[0056] The update gate is divided into two parts, one part decides which new information needs to be updated, and the other part creates a candidate value vector to be added to the cell state. Then, the old cell state is multiplied by the output of the forget gate to discard the part decided to be forgotten, and the new cell state is composed of the multiplication result of the update gate and the candidate value vector. This cell state is like a conveyor belt, on which information can flow smoothly with only a small amount of linear interaction, ensuring the effective transmission of long-term memory. Finally, the output gate determines what the hidden state to be output at the current time based on the updated cell state. The entire calculation process can be accurately described by the following set of formulas: ; in this set of formulas, the meanings of various symbols are as follows: and represent the hidden layer outputs at the previous and current times, respectively; input representing the current time step; , , are the activation values of the forget gate, update gate and output gate, respectively; and represent the cell state at the previous time step and the current time step, respectively; is the candidate value vector for updating the cell state; , , , is the weight matrix of the corresponding gating unit; , , , is the corresponding bias vector; represents the Sigmoid activation function, whose output range is between (0, 1), used for gating signals; represents the hyperbolic tangent activation function, whose output range is between (-1, 1); represents element-wise multiplication.

[0057] In the framework of the present embodiment, the LSTM receives the feature vector with high information content that has been deeply processed by the BiTCN and TPA, which enables it to focus more on learning the long-term temporal evolution rules between these key features, thereby making accurate predictions. Optionally, in order to generate the final single-point or multi-point prediction value, the output of the LSTM module can be connected with one or more fully connected layers for mapping the high-dimensional hidden state to the final prediction dimension.

[0058] In order to objectively evaluate the effectiveness and superiority of the method proposed in the present embodiment, i.e. step S4, we conducted a series of comparative experiments. In the experiment, three widely recognized evaluation indicators were used to measure the prediction effect of the model, namely the root mean square error RMSE, the mean absolute percentage error MAPE and the mean absolute error MAE. Their calculation formulas are as follows: ; wherein, is the total number of prediction points, and represent the predicted data and the true data at the th load point, respectively.

[0059] We compared the VMD-BiTCN-LSTM model proposed in this embodiment with multiple benchmark models, including the traditional BP neural network, the single LSTM model, and the VMD-LSTM and VMD-BiTCN combined models. All models were trained and evaluated on the same training set and test set. The experimental results are clearly recorded in Table 1 below: Table 1 Comparison of evaluation indicators of each model From the data in Table 1, it can be concluded that the improved VMD-BiTCN-LSTM model proposed in this embodiment has significantly reduced all three evaluation indicators compared to all other comparison models, indicating that the model has the optimal performance advantage in terms of prediction accuracy and error control. For example, the MAPE value of the model is only 0.1537, which is much lower than the 0.5127 of the single LSTM model and the 0.4409 of the simple combined VMD-LSTM. At the same time, referring to Figure 6 , the figure shows the comparison of the prediction curve and the true value curve of each model in a typical period of time. From the figure, it can be seen that the fitting degree of the prediction curve of the model and the true load curve (red line) is the highest among all models, especially in the peak and valley regions of the load, its followability and accuracy perform particularly well. The prediction curves of other comparison models have more or less obvious deviations, lags, or problems of insufficient capture of fluctuations. These quantitative data and visual results jointly prove the effectiveness and superiority of the method .

[0060] In summary, the short-term power load forecasting method elaborated in this embodiment combines the nonlinear feature screening based on Copula entropy, the fine signal decomposition based on VMD, and a collaborative prediction model deeply integrating the advantages of BiTCN, TPA, and LSTM together, to construct a complete, efficient, and outstanding performance end-to-end prediction process. This method not only theoretically solves the limitations of existing technologies in multiple aspects, but also proves through experimental verification on real data sets that it can achieve leading prediction accuracy and stability in practical applications, and has very high engineering application value and promotion prospects.

[0061] Embodiment Two: The embodiment provides a short-term power load forecasting system, which is a physical or logical carrier of the method described in embodiment one and is configured to perform all or part of the steps of the foregoing method to achieve high-precision power load forecasting. The system can be implemented as a centralized server, a high-performance workstation, a dedicated embedded device, or a distributed computing cluster composed of multiple computing nodes. Regardless of its physical form, from a functional structure, the system contains several core modules connected to each other and working cooperatively.

[0062] Specifically, the short-term power load forecasting system at least includes a feature screening module, a sequence processing module, and a prediction module, which are connected through explicit data interfaces or buses to ensure smooth data flow.

[0063] The feature screening module is the "data vanguard" of the system, and its core function is to perform the Copula entropy-based feature screening method described in step S1 of embodiment one. The module is configured to receive an original data set containing a historical power load sequence and a plurality of original feature sequences, and is further configured to be able to perform the complete process of calculating a Copula density function, calculating a Copula entropy, and sorting and screening a target feature sequence according to the Copula entropy value. The output of the module is a target feature sequence that has been reduced in dimension and contains only core features that have strong nonlinear correlation with the load.

[0064] The input end of the sequence processing module is connected to the output end of the aforementioned feature screening module, and it also receives the original historical power load sequence as another input. The core function of the module is to perform the VMD decomposition and data fusion described in step S2 of embodiment one. It is configured to call the variational mode decomposition VMD algorithm to decompose the input non-stationary historical power load sequence into a plurality of relatively stationary and physically meaningful intrinsic mode function IMF sub-sequences. Subsequently, the module will be configured to align and merge the decomposed IMF sub-sequences in time with the target feature sequence received from the feature screening module, thereby generating a structured, multi-dimensional fusion data set. This fusion data set will be the final input of the system prediction core. In addition, the module can also optionally integrate data preprocessing functions, such as performing the missing value filling, outlier correction, and data normalization operations described in step S21 of embodiment one.

[0065] The prediction module is responsible for performing the most core prediction task. Its input end is firmly connected to the output end of the sequence processing module for receiving the carefully processed fusion data set. The internal structure of the prediction module is designed as a deep series, hierarchical architecture to strictly correspond to the execution order of steps S3 to S5 in embodiment one.

[0066] Specifically, the prediction module internally comprises, in sequence: a bidirectional temporal convolutional network (BiTCN) submodule, a temporal attention (TPA) submodule, and a long short-term memory (LSTM) submodule.

[0067] The flow path of data within the prediction module is determined as follows: the fused dataset first enters the BiTCN submodule, where bidirectional and parallel local dynamic feature extraction is performed; the output end of the BiTCN submodule is directly connected to the input end of the TPA submodule, where the extracted dynamic features are subjected to attention weighting to focus on key periodic information; the output end of the TPA submodule is connected to the input end of the LSTM submodule, where long-term temporal dependency learning is performed based on the weighted high-level features, and a prediction result is finally generated. In order to convert the high-dimensional hidden state output by the LSTM network into a specific load prediction value, the prediction module can further optionally comprise a fully connected layer submodule, the input end of which is connected to the output end of the LSTM submodule, and the output thereof is the final prediction result of the entire system.

[0068] In terms of physical implementation, the system of the present embodiment generally relies on a general-purpose or special-purpose computer device. The computer device comprises at least one or more processors, such as a central processing unit (CPU), a graphics processing unit (GPU), or a tensor processing unit (TPU), and a memory, such as a random access memory (RAM), a solid state drive (SSD), or a hard disk drive (HDD), connected thereto.

[0069] The aforementioned feature screening module, sequence processing module, prediction module, and various submodules thereof can all be computer executable instructions, i.e., computer programs, which are solidified or loaded into the memory. When the processor reads and executes these computer programs stored in the memory, it will drive the hardware to complete all the method steps described in Embodiment I in sequence, so that the entire computer device operates as a functionally complete short-term power load prediction system.

[0070] In addition, the system can also include a human-computer interaction interface for data input and result output, and a network communication interface for data exchange with external data sources, such as a meteorological database and a power grid SCADA system.

[0071] In summary, the short-term power load prediction system disclosed in the embodiment provides a complete, reliable and implementable hardware and software platform for the advanced prediction method proposed in Embodiment I through its modular functional design and clear internal connection relationship. The system can automatically complete the whole process from raw data processing to high-precision load prediction through the integration of the functional modules of Copula entropy screening, VMD decomposition and BiTCN-TPA-LSTM composite model. The objective technical effect lies in providing an entity device or logic system that can stably and efficiently output more accurate short-term power load prediction results than the prior art, thereby playing a significant application value in actual power grid dispatching and management.

Claims

1. A method for electric load forecasting by fusing Copula entropy and VMD-BiTCN-LSTM, characterized in that, The method comprises the following steps: S1: obtaining a dataset comprising a historical power load sequence and a plurality of original feature sequences, and using a feature screening method based on Copula entropy to screen a target feature sequence with strong correlation with the historical power load sequence from the plurality of original feature sequences; S2: decomposing the historical power load sequence into a plurality of intrinsic mode function (IMF) sub-sequences using a variational mode decomposition (VMD) algorithm, and fusing the plurality of IMF sub-sequences with the target feature sequence screened in step S1 to obtain a fused dataset; S3: performing bidirectional feature extraction on the fused dataset using a bidirectional time convolutional network (BiTCN) to obtain dynamic features; S4: weighting the dynamic features using a time attention mechanism (TPA) to enhance load cycle characteristics to obtain weighted features; S5: inputting the weighted features into a long short-term memory (LSTM) network for time series learning to generate a short-term power load prediction result.

2. The method of claim 1, wherein the method is characterized by, The feature screening method based on Copula entropy in step S1 specifically comprises: S11: For any original feature sequence and the historical power load sequence, calculate its Copula density function ; S12: Calculate the corresponding Copula entropy according to the Copula density function ; S13: selecting an original feature sequence with a Copula entropy less than a preset threshold as the target feature sequence according to the principle that the smaller the Copula entropy, the stronger the correlation.

3. The method of claim 1, wherein the method is characterized by, After step S2 and before step S3, the method further comprises: S21: preprocessing the fused dataset, wherein the preprocessing comprises linear interpolation filling of missing values, correction of abnormal values, and data normalization.

4. The method of claim 1, wherein the method is characterized by, After step S5, the method further comprises: mapping the output of the LSTM network through a fully connected layer to generate the short-term power load prediction result.

5. The method of claim 1, wherein the method is characterized by: In step S3, the BiTCN divides and processes the fused dataset in parallel using a sliding window.

6. The method of claim 1, wherein the method is characterized by, The internal calculation process of the LSTM network in step S5 comprises: In the formula, , are the hidden layer outputs at the previous time and the current time, respectively. t for the current time instant; , , are weights of the forget gate, update gate, and output gate, respectively; is a vector of candidate values; , Ci-1 and Ci are the cell states of the previous and current time, respectively. with is an activation function.

7. A system for electric load forecasting fusing Copula entropy and VMD-BiTCN-LSTM, configured to perform the method of claims 1-6, characterized in that, The method comprises: a feature screening module configured to perform steps S11 to S13 of claim 2 to screen a target feature sequence from a dataset comprising a historical power load sequence and a plurality of original feature sequences; a sequence processing module connected to the output end of the feature screening module and configured to decompose the historical power load sequence into a plurality of intrinsic mode function (IMF) sub-sequences using a variational mode decomposition (VMD) algorithm, and fuse the IMF sub-sequences with the target feature sequence to generate a fused dataset; a prediction module connected to the output end of the sequence processing module and configured to receive the fused dataset and generate a short-term power load prediction result; The prediction module comprises, in sequence, a bidirectional time convolutional network (BiTCN) submodule for extracting dynamic features, a time attention mechanism (TPA) submodule connected to the output end of the BiTCN submodule for weighting the dynamic features, and a long short-term memory (LSTM) submodule connected to the output end of the TPA submodule for generating a final prediction result. ​ ​ ​ 8. The system for electricity load forecasting fusing Copula entropy and VMD-BiTCN-LSTM of claim 7, wherein: The prediction module further comprises a full connection layer submodule, an input end of the full connection layer submodule being connected to an output end of the long short-term memory network LSTM submodule, and the full connection layer submodule being configured to map the output of the LSTM submodule. 9.A computer device, comprising a processor and a memory, wherein the memory has stored thereon a computer program, and the computer device is characterized in that, The processor executes the computer program to implement the method in any one of claims 1-6.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method in any one of claims 1-6.

Images