Power load prediction method based on bidirectional attention time sequence convolutional network and probability mode correction

Abstract

The invention relates to the technical field of power load energy prediction, and provides a prediction method based on a bidirectional attention time sequence convolutional network and a probability mode correction model. Aiming at the limitation of an existing traditional model in the aspects of long-sequence dependence modeling and feature dynamic weighting and the problem that it is difficult to deeply analyze the internal structure of power load data and optimize features, the invention proposes that on the basis of feature extraction of a bidirectional time sequence convolutional network (BiTCN), a channel attention mechanism (CAM) is fused, and a power load model is obtained. Adaptive weighting is carried out on the extracted features so as to automatically identify key features; meanwhile, a probability mode correction mechanism is introduced, and data points are reasonably divided, so that the depicting capability of a complex load mode is enhanced. The method provided by the invention significantly improves the precision and robustness of power load prediction, is suitable for power grid dispatching and energy optimization management, and provides reliable technical support for efficient utilization of power energy.

Description

Technical fields:

[0001] The embodiments of the present invention relate to the field of time series forecasting, and more specifically, the embodiments of the present invention relate to a power load forecasting method based on time series feature extraction and probabilistic pattern correction. Background technology:

[0002] In the field of power load forecasting, existing technologies mainly include three types of methods: traditional time series models (such as ARIMA), shallow machine learning models (such as SVM and random forest), and classic deep learning models (such as RNN, LSTM, and CNN). These methods have achieved certain results in their respective suitable application scenarios, but they still have the following limitations:

[0003] First, while traditional time series models can capture linear trends, they are not good at modeling nonlinear features; shallow machine learning models are susceptible to feature redundancy and have limited generalization ability; classic deep learning models suffer from gradient explosion or vanishing problems, difficulty in capturing long-term dependencies, and excessive computational complexity.

[0004] Secondly, existing methods generally suffer from three defects and limitations: first, the feature weights are fixed and cannot adapt to the dynamic changes in feature importance; second, time series modeling is mostly unidirectional and cannot utilize historical and future information simultaneously; and third, there is a lack of effective feature correction mechanisms, making the model sensitive to noise and outliers.

[0005] To address the aforementioned issues, this invention proposes a power load forecasting model based on a Bi-Temporal Convolutional Network (Bi-TCN), Channel Attention Mechanism (CAM), and Probabilistic Pattern Correction. This model extracts forward and backward temporal features in parallel using Bi-TCN to fully capture long-term dependencies; it utilizes CAM to dynamically learn channel weights and adaptively strengthen key features; and it combines DBSCAN clustering and a probabilistic classifier based on a Gaussian mixture model to statistically model and correct the features, effectively suppressing noise and outlier interference. Finally, a GMDH network is used to achieve feature fusion and prediction, constructing a power load forecasting model that balances accuracy and robustness. Summary of the Invention:

[0006] This invention provides the following technical solution:

[0007] The method described in this invention includes a three-level progressive intelligent prediction process, specifically comprising the following steps:

[0008] Historical power load data (multidimensional feature data) is acquired and preprocessed (data cleaning, missing value imputation and Z-score standardization to obtain standardized time series data to eliminate the influence of different feature dimensions on model training).

[0009] Furthermore, basic temporal features are extracted through a bidirectional temporal convolutional network, wherein the temporal network employs left-padded causal dilation convolution to ensure temporal dependencies; the formula for left-padded causal dilation convolution is as follows:

[0010]

[0011] Where F(t) is the output sequence after convolution at time t, K is the size of the convolution kernel, f(k) is the weight parameter of the filter at position k, d is the dilation factor, X is the historical input data sequence, b is the bias term of the convolution operation, and left padding is achieved by precisely padding (K-1)×d zero values ​​at the beginning of the input sequence.

[0012] Furthermore, each convolutional layer introduces a residual connection structure to build deeper networks and alleviate the gradient vanishing problem. Its output formula is as follows:

[0013] Output=Activation(F(t)+Wres·Input)

[0014] Where F(t) is the output of the causal dilation convolution operation in this layer; Input is the input data of the residual block; Wres is the weight matrix of a 1x1 convolution.

[0015] The basic temporal features are further processed using a channel attention mechanism with channel weighting. The formula for calculating the channel attention weight matrix Mc is as follows:

[0016] Mc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0017] Where F is the input feature map with dimensions (BatchSize, Channels, TimeSteps), σ represents the Sigmoid activation function, MLP represents the shared multilayer perceptron, AvgPool represents the global average pooling function, and MaxPool represents the global max pooling function.

[0018] Furthermore, an unsupervised learning method based on density clustering is adopted to perform pattern segmentation on the weighted features, and the optimal hyperparameters of the DBSCAN algorithm are determined by combining the elbow rule and the silhouette coefficient.

[0019] Furthermore, the features are statistically modeled and probabilistically corrected using a probabilistic pattern classifier. For the i-th cluster Ci obtained from cluster analysis, the statistical features of all data points within it are calculated, including the mean vector μi and covariance moment Σi. The multivariate Gaussian probability density function P(x|Ci) for this cluster is constructed to calculate the probability that any data point x is generated by this cluster, as shown in the following formula:

[0020]

[0021] Where k is the dimension of the eigenvector, and |Σi| is the determinant of the covariance matrix Σi;

[0022] For a new power load data point, the posterior probability P(Ci|x) of belonging to cluster Ci is calculated according to Bayes' theorem, achieving probabilistic soft classification rather than hard partitioning. The formula is as follows:

[0023]

[0024] Where P(Ci) is the prior probability of cluster Ci, and K is the total number of clusters;

[0025] Feature fusion and optimization of probability weights are performed on the data points. The posterior probability P(Ci|x) of data point x belonging to each cluster is normalized to ensure that the sum of all weights is 1, resulting in the normalized weight Wi. The normalization formula is as follows:

[0026]

[0027] The original features x or attention-weighted features of the data points are weighted and fused to obtain the final probabilistically corrected features Xcorrected:

[0028]

[0029] Where K is the number of clusters.

[0030] Based on the corrected features, the hybrid forecasting model GMDH is used to forecast the load and output the final forecast results.

[0031] The specific advantages are as follows:

[0032] 1. This invention employs a Bi-TCN model to extract temporal features. It fully leverages the advantages of the Bi-TCN model, such as its ability to extract long-term dependencies in both forward and backward directions, efficient parallel computing, and a stable training process. Furthermore, the scalability and flexibility of the Bi-TCN model allow us to adjust and optimize the model according to the needs of different application scenarios, adapting it to a wider range of electricity consumption prediction tasks.

[0033] 2. This invention employs a channel attention mechanism to achieve adaptive allocation of feature weights. By focusing the model's main attention on key feature channels, it effectively improves the utilization rate of feature information and the accuracy of prediction. Simultaneously, attention-weighted features have strong anti-interference capabilities, resisting interference from external factors such as noisy features to a certain extent, thus enhancing the stability and reliability of prediction results.

[0034] 3. This method uses a probabilistic pattern classifier to process feature correction. By using the probabilistic pattern classifier, the input features are effectively grouped and the probability of belonging to each pattern class is calculated, which helps to identify abnormal features and retain key features, thereby improving the efficiency and accuracy of subsequent processing. Attached image description:

[0035] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example, and not limitation, in which:

[0036] Figure 1 A flowchart of the power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction is provided for an example of this method.

[0037] Figure 2 This is a schematic diagram of the data preprocessing workflow in this invention;

[0038] Figure 3 This is a schematic diagram of the workflow of the temporal feature extractor in this invention;

[0039] Figure 4 This is a schematic diagram of the channel attention mechanism in this invention;

[0040] Figure 5 This is a schematic diagram of the workflow of the unsupervised clustering algorithm in this invention;

[0041] Figure 6 This is a schematic diagram of the workflow of the probabilistic pattern classifier in this invention;

[0042] Figure 7 This is a schematic diagram of the final prediction process in this invention;

[0043] Figure 8 A comparison chart showing the prediction results of the method provided in this embodiment of the invention with those of other models on dataset D1;

[0044] Figure 9 A comparison chart showing the prediction results of the method provided in this embodiment of the invention with those of other models on dataset D2. Detailed implementation method:

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. This section aims to provide clear and complete technical guidance to those skilled in the art, ensuring the feasibility of this invention.

[0046] Exemplary methods

[0047] The power load forecasting method based on bidirectional attention temporal convolutional networks and probabilistic pattern correction in this invention includes the following steps: Figure 1 A detailed process according to an embodiment of this disclosure is illustrated schematically.

[0048] Step S1: Data Acquisition and Preprocessing

[0049] This step is fundamental to model training and prediction, and its quality directly determines the final performance. The specific implementation is as follows:

[0050] Obtain historical power load data: Collect multidimensional feature data aligned with its time period to form the original power load time series dataset. Where y t x represents the load value at time t. t This corresponds to the feature vector. It includes meteorological features (temperature, humidity, wind speed, solar radiation intensity, etc.) and social behavioral features (family size, holiday signs, peak hours, etc.).

[0051] Data preprocessing: Transforming raw data into a format suitable for model training and analysis; aiming to improve data quality, model performance, and accuracy, making it more suitable for subsequent analysis and modeling. Data cleaning: Identifying and handling obvious outliers (such as negative load values ​​or values ​​far exceeding reasonable ranges), which can be achieved using quantile-based outlier detection methods; Missing value imputation: For continuous features (such as temperature, load), time-series sensitive methods are used for imputation, such as linear interpolation or using the previous valid value (Forward Fill). For categorical features, the mode is used for imputation; Z-score standardization: To eliminate the influence of dimensions, each feature dimension is standardized. For feature f, its standardized values ​​are as follows:

[0052]

[0053] Where: μ f and σ f These are the mean and standard deviation of the feature on the training set, respectively.

[0054] Step S2: Bidirectional Temporal Feature Extraction

[0055] The core of this module is to construct a bidirectional temporal convolutional neural network (Bi-TCN) to process standardized temporal data X∈R. B×C×L (B represents batch size, C represents the number of feature channels, and L represents sequence length) Extract deep bidirectional features. Forward path: Receive input sequence X, and process it in ascending time order (x1, x2, ..., x...). L Processing; Reverse path: First, reverse the input sequence along the time axis to obtain X. reverse =(x L,x L-1 ,......,x1), and then process them in this order. This operation aims to uncover potential dependencies that allow us to infer the current state from the "future" state.

[0056] Left-padded causal dilated convolution: Both pathways use multi-layer causal dilated convolutions as basic units. Its mathematical formalization is as follows:

[0057]

[0058] The key to its implementation lies in left padding. To achieve causality and maintain sequence length, (K-1)×d zeros need to be precisely padded to the left of the input sequence; for example, when the kernel size K=3 and the inflation factor d=2, the padding number is (3-1)×2=4; this ensures that when calculating the output at time t, the receptive field of the kernel only covers the input at times t, t-2, and t-4, strictly following the causal law.

[0059] Inflation factor setting: To effectively capture long-range dependencies, a larger inflation factor is used in deeper layers of the network; for the i-th layer, the inflation factor d = 2. i-1 This exponential growth strategy causes the receptive field to expand exponentially with layer depth. Residual connections: To prevent gradient vanishing in deep networks and preserve shallow features, residual connections are introduced within each convolutional layer (or residual block). The output is as follows:

[0060] Output = Activation(F(t) + W) res Input)

[0061] In the formula: F(t) is the output after convolution, weight normalization, and ReLU activation function; W res It is a 1×1 convolution;

[0062] Bidirectional feature fusion: converting the forward path output H forward ∈R B×C'×L With reverse path output H reverse ∈R B×C'×L Concatenation is performed along the feature channel dimension:

[0063] H fused =Concat(H forward H reverse )∈R B×2C'×L

[0064] This feature is fused with H fused It also incorporates historical context information and hints of future trends, forming the input for subsequent modules—the initial fusion features.

[0065] Step S3: Dynamic Feature Enhancement

[0066] This module introduces a Channel Attention (CAM) mechanism to adaptively recalibrate the initial fused features output by Bi-TCN, such as... Figure 3 As shown.

[0067] Global context description: For the input feature map F∈R B×2C'×L Global average pooling (GAP) is performed along the time step dimension: Z avg =GAP(F), where It captures the global average response of the features;

[0068] Global Max Pooling (GMP): Z max =GMP(F), where It captures the most prominent and discriminative information among the features.

[0069] Shared Multilayer Perceptron (MLP): [The text abruptly ends here, likely due to an incomplete sentence or a avg and Z max Each input is a shared MLP consisting of two fully connected layers. The MLP first reduces the number of channels from C to C / r (where r is the reduction ratio) using the ReLU activation function; then it increases the dimensionality back to C. This process is represented as follows:

[0070] s avg =W2·δ(W1·z avg ),s max =W2·δ(W1·z max )

[0071] In the formula: W1∈R (C / r)×C W2∈R C×(C / r) δ is the activation function.

[0072] Attention weight generation: The two output vectors of the MLP are added element-wise, and the final channel attention weight matrix M is obtained by passing it through the Sigmoid activation function. c ∈R B×C×1

[0073] M c =σ(S avg +S max )

[0074] In the formula: σ is the Sigmoid function, which compresses the weight values ​​to the interval [0,1].

[0075] Feature recalibration: The attention weights are multiplied channel-wise with the original input feature map as follows:

[0076]

[0077] In the formula: ⊙ represents channel-by-channel broadcast multiplication.

[0078] Step S4: Analysis of the intrinsic structure of the data

[0079] This module uses the DBSCAN density clustering algorithm to discover typical operating patterns hidden in power load data under unsupervised conditions, such as... Figure 4 As shown.

[0080] Feature preparation: Weighted features output from the CAM module (For the entire training dataset) Reshape it into a two-dimensional matrix, with each row representing a feature vector at a time point.

[0081] Parameter determination: DBSCAN has two key hyperparameters: neighborhood radius Eps and minimum number of core points MinPts. Elbow algorithm: Calculate the number of core points corresponding to different Eps values ​​and plot the curve. Select the Eps value at the inflection point of the curve.

[0082] Silhouette coefficient: For different combinations of (Eps, MinPts), run DBSCAN and calculate the average silhouette coefficient of all non-noise points. The closer the silhouette coefficient is to 1, the denser the clusters are and the better the separation between clusters. Combining both factors, select the parameter combination that results in a high silhouette coefficient and clustering results with clear physical meaning (e.g., corresponding to base load, weekday peak, weekend patterns, etc.).

[0083] Clustering execution: Run the DBSCAN algorithm. This algorithm, based on density reachability and connectivity, groups high-density regions in the feature space that are density-connected into clusters, and can discover clusters of arbitrary shapes. Data points in low-density regions that cannot be assigned to any cluster are marked as noise points (-1); finally, K clusters {C1, C2, ..., C...} are obtained. K} and a set of noise points.

[0084] Step S5: Probabilistic Pattern Modeling

[0085] This module constructs a probabilistic pattern classifier based on a Gaussian mixture model (GMM) for each cluster obtained in S4.

[0086] Statistical feature calculation: For the i-th cluster C i Calculate the mean vector and covariance matrix of all data points within it.

[0087] Mean vector μ i : Indicates the central location of this load pattern in the feature space;

[0088] covariance matrix Σ i This describes the distribution shape of data points within the cluster and the correlation between different feature dimensions. Probability density function construction: based on μ. i and Σ i Construct the multivariate Gaussian probability density function for this cluster:

[0089]

[0090] In the formula: k is the dimension of the feature vector.

[0091] Posterior probability calculation: For a new data point x, calculate its membership in cluster C according to Bayes' theorem. i Posterior probability:

[0092]

[0093] Wherein, the prior probability P(C) i The probability can be estimated by the proportion of data points within that cluster to the total data volume (excluding noise points). The denominator is a normalization factor to ensure that the sum of all posterior probabilities is 1.

[0094] Step S6: Feature Probability Correction

[0095] This module uses posterior probabilities to perform soft correction on features, enhancing the model's ability to handle uncertainties and mixed patterns.

[0096] Probability normalization: To ensure the sum of the weights is 1, the posterior probability is normalized. In practice, since the posterior probability itself is already normalized using Bayes' theorem, this step can be omitted. However, to ensure robustness, it can be normalized again:

[0097]

[0098] Weighted feature fusion: utilizing the normalized weights W i As a confidence level, the attention-weighted features of the data points are weighted and fused to obtain the final features after probability correction:

[0099]

[0100] The advantage of this mechanism is that for data points located at cluster boundaries or in areas of pattern overlap, the final feature representation is a weighted average of the features of all possible patterns, rather than being rigidly divided into a single pattern. This achieves a smooth transition in feature representation and significantly improves the robustness of the model.

[0101] Step S7: Load Forecast Output

[0102] This module will convert the probability-corrected final feature X... corrected This is mapped to the final load forecast.

[0103] Regression prediction layer: Typically consists of one or more fully connected layers. The last fully connected layer outputs a single neuron, corresponding to the predicted power load at a specific future point in time.

[0104]

[0105] In the formula: W o and b o These are the weights and biases for the output layer.

[0106] Model training: The entire model (from BiTCN to the regression layer) is trained in an end-to-end manner.

[0107] Predictive output: After training, the model can receive new time series data and, through the complete process from S1 to S7, output the predicted value of future power load, providing a direct and reliable basis for intelligent scheduling and decision management of the power system.

[0108] Preferred embodiment

[0109] This study uses two power load datasets for validation: the first is "electric" and the second is "household_energy_consumption"; the two datasets are named D1 and D2. This study conducts in-depth analysis and research on the two datasets.

[0110] D1: Data was collected from 00:00 on December 11, 2017 to 17:00 on June 25, 2020, on an hourly basis. It contains 22,200 records, covering fifteen characteristic variables including daily maximum temperature (Homestead_maxtempC), daily minimum temperature (Homestead_mintempC), and dew point temperature (Homestead_DewPointC), with consumption added as an output variable for further analysis.

[0111] D2: Data collection covered the period from April 1st to April 7th, 2025, totaling 12,857 household data points. This included 90,001 records, encompassing four characteristic variables: household size (Household_Size), average temperature (Avg_Temperature_C), air conditioning ownership (Has_AC), and peak hour electricity consumption (Peak_Hours_Usage_kWh). Load consumption (Energy_Consumption_kWh) was also included as an output variable for further analysis.

[0112] The development and implementation of the experiments were carried out using the PyCharm integrated development environment and deep learning frameworks such as TensorFlow. The data required for the experiments was obtained from CSV files, and then preprocessed, including but not limited to normalization, denormalization, and the application of the sliding window technique. In the early stages of model training, the Adam optimizer was used to optimize the model parameters, accelerating model convergence and improving generalization ability. To prevent the model from losing generalization ability due to overfitting the training data, an early stopping strategy was implemented, terminating training promptly when the performance on the validation set no longer improved, ensuring efficient and robust model training.

[0113] Prediction Results Analysis

[0114] In this study, three basic models (CNN, LSTM, RNN) and six hybrid models (CNN+LSTM, TCN+LSTM, TCN+Transformer, DWT+Transformer, BiGRU+BiTCN, VMD+BiLSTM) were selected for comparative experiments.

[0115] The experimental results are shown in Table 1. It can be seen that BiTCN+CAM+PP exhibits optimal performance in the test scenarios of both datasets. Taking the prediction results of D1 as an example, compared with CNN, LSTM, RNN, CNN+LSTM, TCN+LSTM, TCN+Transformer, DWT+Transformer, BiGRU+BiTCN, and VMD+BiLSTM, the MSE of our proposed method decreases by 35.78%, 36.68%, 42.19%, 28.15%, 21.79%, 35.92%, 9.29%, 10.81%, and 6.19%, respectively. Taking the prediction results of D2 as an example, the evaluation metric R of our model... 2 The value is 0.9876, making it the only model to reach this level (the remaining models range from 0.982 to 0.986). The above results indicate that this model performs better than other models that only capture time-domain information, showing a significant improvement in prediction accuracy and possessing potential practical application value.

[0116] Table 1. Evaluation metrics for the dataset comparison experiment.

[0117]

[0118] To more accurately quantify and intuitively illustrate the gap between predicted results and actual observations, this paper specifically plots... Figure 7 and Figure 8These two graphs respectively show the changes in the predicted and actual values ​​of D1 and D2, providing a visual representation of the deviation between the actual and predicted values. By comparing the graphs, it can be found that compared with other models, the BiTCN+CAM+PP model has a significantly smaller error between the predicted and actual values ​​when predicting the electricity load of D1 and D2. This finding is highly consistent with the data presented in Table 1, further verifying the superior performance of the BiTCN+CAM+PP model in electricity load prediction.

[0119] This embodiment proposes a power load forecasting method based on bidirectional attention temporal convolutional networks and probabilistic pattern correction for the task of accurate prediction of power energy consumption. A power load forecasting model is established, integrating three techniques: bidirectional temporal feature extraction, channel attention weighting, and probabilistic pattern correction. An end-to-end training method combining unsupervised clustering and probabilistic correction is designed to improve the model's generalization ability. Experimental results on two different publicly available power load datasets demonstrate the superiority of this embodiment. Furthermore, when the training data contains noise and fluctuations, the experimental results of this embodiment show significant improvements over other traditional and hybrid models, exhibiting more robust performance. The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in this invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of this invention.

Claims

1. A power load forecasting method based on bidirectional attention temporal convolutional networks and probabilistic pattern correction, characterized in that, Includes the following steps: Historical power load data is acquired and standardized preprocessed. Basic temporal features are extracted using a bidirectional temporal convolutional network, wherein causal dilated convolution is used to ensure temporal dependencies. Channel attention mechanism is used to perform channel weighting on the basic temporal features. Unsupervised learning method based on density clustering is used to perform pattern segmentation on the weighted features. Statistical modeling and probability correction of the features are performed using a probabilistic pattern classifier. Based on the corrected features, a hybrid prediction model is used to predict the load and output the final result.

2. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The data acquisition and preprocessing module includes: data cleaning, missing value imputation, and Z-score standardization of the original dataset features to eliminate the influence of different feature scales on model training.

3. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The bidirectional temporal convolutional network includes a combination structure of a forward temporal convolutional network and a backward temporal convolutional network; left-filled causal dilation convolution is used to ensure the correctness of temporal dependencies; and residual connection mechanism is used to alleviate the gradient vanishing problem in deep networks.

4. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The channel attention mechanism includes: receiving the extracted basic temporal features as input, performing global average pooling and global max pooling operations to obtain channel information; learning the inter-channel dependencies through a two-layer fully connected network, where the first layer uses the ReLU activation function and the second layer uses the Sigmoid activation function, and a Dropout layer is introduced between the two fully connected network layers for regularization; and multiplying the temporal feature vector with the learned channel attention weights channel by channel to achieve feature weighting.

5. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The unsupervised learning method includes: using the elbow algorithm to determine the optimal clustering parameter range; using the DBSCAN density clustering algorithm to partition the feature space, with the parameters configured as a neighborhood radius of 0.5 and a minimum number of samples of 5; implementing a noise point identification and suppression mechanism to handle abnormal data; calculating the silhouette coefficient to evaluate the clustering quality and selecting the optimal clustering result; and outputting clustering labels for subsequent probabilistic pattern classification.

6. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The probabilistic pattern classifier includes: performing feature statistics on each cluster, calculating the mean vector and covariance matrix; constructing a probability density function for each cluster based on a multivariate Gaussian distribution; calculating the posterior probability of a feature sample belonging to each cluster to form a probability weight matrix; making a classification decision based on the principle of maximizing the posterior probability; and using the probability weights to perform weighted correction on the original features to output optimized features.

7. The power load forecasting method based on bidirectional attention temporal convolutional network and probabilistic pattern correction according to claim 1, characterized in that, The hybrid prediction model includes: using a hybrid GMDH model to make the final prediction for the corrected features; outputting the load prediction value and calculating multi-dimensional evaluation indicators, including mean square error, root mean square error, coefficient of determination, mean absolute error and correlation coefficient; outputting the prediction results and saving them.

Images