An ultra-short-term power load prediction method and device

Abstract

The present application relates to the technical field of power load prediction, and particularly relates to a kind of ultra-short-term power load prediction method.The method includes three parts of time series decomposition, reconstruction of the sequence decomposed based on sample entropy, and prediction of the reconstructed sequence using neural network.CEEMDAN is used to decompose time series, CEEMDAN can adapt to the local characteristics of data, automatically extract the internal oscillation mode from complex signals, and the IMFs obtained by decomposition can be directly used for feature engineering to support the establishment of more accurate prediction model.The ultra-short-term load prediction model uses TSFormer neural network, and the TSFormer architecture can effectively capture the long-term dependencies in time series data through self-attention mechanism, which is particularly critical for accurate power load prediction.Compared with traditional RNN and LSTM, Transformer can process all time steps in parallel, improving training efficiency and prediction performance.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting, and in particular to an ultra-short-term power load forecasting method based on Complete Ensemble Empirical Mode Decomposition (CEEMDAN) and TSFormer neural network. Background Art

[0002] Power load forecasting is fundamental to grid operation management, dispatching, and planning. With the adjustment of industrial structures and the implementation of sustainable development policies, load forecasting for large industrial users has become a key focus for power supply companies. In the management and optimization of industrial park power systems, accurate load forecasting is crucial for ensuring power supply stability, optimizing power resource allocation, and reducing energy costs. Ultra-short-term load forecasting (typically within hours or even a single day) has become particularly critical, especially with the increasing liberalization of power markets and the increasing complexity of power grids. Effective ultra-short-term load forecasting can help power companies dynamically adjust power generation and procurement strategies, improving grid responsiveness and economic efficiency. Time series analysis (such as ARIMA models): These models are widely used for load forecasting and rely on historical data to predict future load. However, they primarily apply to linear patterns in data and have limited effectiveness in handling nonlinear and non-stationary data. They are unable to effectively respond to sudden changes and non-periodic fluctuations in power load data, particularly during extreme weather or emergencies, where forecast accuracy can be significantly reduced. Regarding support vector machines (SVMs), decision trees, and random forests: These methods can handle nonlinear problems, but typically require extensive data preprocessing and feature selection, and model training and parameter tuning are complex and time-consuming. The disadvantage is that they require a large amount of labeled data and have poor adaptability to unknown patterns. In ultra-short-term forecasting scenarios, these models are often not suitable due to their slow response time.

[0003] With the recent rise of deep learning, many deep learning-related algorithms have significantly improved the accuracy of short-term power load forecasting. While traditional neural networks and deep learning networks can learn complex patterns in data, they are inadequate for capturing the inherent temporal dependencies of time series data. However, when processing time series data, balancing long-term memory and short-term input is difficult, leading to the potential for overlooking important historical information or failing to effectively respond to recent data changes in practical applications. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an ultra-short-term power load forecasting method based on CEEMDAN and TSFormer neural networks, which captures the short-term and long-term time dependencies of ultra-short-term data, enables the forecast results to accurately reflect the load change trend, and is used to improve the accuracy, adaptability and real-time performance of the forecast.

[0005] In order to solve the above technical problems, the present invention adopts a technical solution: a method for ultra-short-term power load forecasting, comprising the following steps:

[0006] S01) Decompose the original load sequence using the synchronous extension of the fully integrated empirical mode decomposition, decompose multiple IMF component sequences from the nonlinear and non-stationary power data, and extract the spatiotemporal features synchronously;

[0007] S02), reconstructing the IMF component sequence obtained by decomposition in step S01) based on sample entropy to obtain two new sequences, namely the fluctuation sequence F1 and the trend sequence F2;

[0008] S03) Input the fluctuation sequence F1 and the trend sequence F2 into the short-term load forecasting model composed of the TSFormer network. The TSFormer network includes an input layer, an encoder layer, and an output layer. The input layer performs feature embedding and position encoding on the input sequence. Feature embedding converts the input sequence into an embedding vector in a high-dimensional space, and position encoding adds position encoding to the embedding vector to retain time information. The encoder layer is composed of multiple stacked sub-encoding layers, each of which includes a self-attention layer and a feedforward network. The self-attention layer transforms the position-encoded embedding vector through different query, key, and value matrices, and then performs scaled dot product attention calculation to obtain the final attention output. Each self-attention layer is followed by a feedforward network, which independently applies the same fully connected layer to the output of each position for linear transformation and activation. The output layer is located after the encoder layer and performs the final prediction through one or more fully connected layers to predict the power load values ​​at the next few time points or a certain moment.

[0009] Furthermore, the self-attention layer adopts a multi-head self-attention mechanism, that is, the position-encoded embedding vector is transformed in parallel on multiple sets of different query, key, and value matrices, and the output results are concatenated and linearly transformed again to obtain the final attention output.

[0010] Furthermore, feature embedding is implemented through a fully connected layer, mapping the features of each time point to a vector of fixed size.

[0011] Furthermore, the positional encoding is learned via a static cosine function.

[0012] Furthermore, the output of each sub-coding layer is added to the input of the layer and then normalized and connected with the residual.

[0013] Furthermore, the feedforward network uses two layers of linear transformation with a ReLU activation function in the middle.

[0014] Furthermore, step S01) is specifically as follows:

[0015] S11) Add Gaussian noise. Add a limited number of adaptive Gaussian white noises to the original load sequence. The formula is:

[0016] (1);

[0017] in represents the original load sequence, represents the sequence obtained by adding Gaussian white noise to the original load sequence for the i-th time, represents the noise figure, represents the Gaussian white noise added for the i-th time;

[0018] S12) Find the first IMF component and use the average value of the first IMF component to find the Perform EMD decomposition as shown in formula (2):

[0019] (2),

[0020] in is the first IMF obtained after decomposing the original load sequence after adding Gaussian white noise for the i-th time, M is the number of Gaussian white noise decompositions, and then the original load sequence is used Subtracting IMF1, we get the first residual sequence, as shown in formula (3):

[0021] (3);

[0022] in is the obtained residual sequence;

[0023] S13) Find the second IMF, Perform EMD decomposition to obtain the second IMF component, and the result is shown in formula (4):

[0024] (4),

[0025] in 、 They represent the operation of performing the j-th EMD decomposition on the corresponding signal and extracting the j-th IMF components. Represents the weight coefficient after adding Gaussian white noise;

[0026] S14) Calculate the final sequence of residuals and repeat formula (5) to obtain the remaining IMF components.

[0027] (5);

[0028] Where k is the total number of IMF components, 、 denotes the kth, k-1th residual sequence, is the kth IMF component, denotes the weight coefficient related to the kth operation, denotes the EMD decomposition of the corresponding signal, respectively.

[0029] The final residual sequence is (6), wherein res is the final residual sequence.

[0030] Further, the input of the short-term load forecasting model also includes weather data, date type.

[0031] Further, the decomposition result of step S01) includes a plurality of IMF component sequences and a residual, each IMF component sequence corresponding to a specific frequency component in the original load signal.

[0032] An ultra-short-term power load forecasting device, comprising a processor and a memory storing program instructions, the processor being configured to execute the SEEMDAN and TSFormer neural network-based ultra-short-term power load forecasting method as described above when running the program instructions.

[0033] The beneficial effects of the present application are:

[0034] 1. CEEMDAN can adapt to the local characteristics of data and automatically extract the internal oscillation mode (called intrinsic mode function, IMF) from complex signals, which is crucial for nonlinear and non-stationary data analysis.

[0035] 2. The IMFs obtained by decomposition can be directly used for feature engineering, supporting the establishment of more accurate prediction models. In power load forecasting, different IMFs can correspond to different load fluctuation patterns, which helps to build prediction models reflecting these patterns.

[0036] 3. The TSFormer architecture can effectively capture long-term dependencies in time series data through self-attention mechanisms, which is particularly crucial for accurate power load forecasting. Compared with traditional RNN and LSTM, Transformer can process all time steps in parallel, improving training efficiency and prediction performance.

[0037] 4. Due to the design of its self-attention layer, the TSFormer model can achieve high parallelization when processing sequence data, significantly reducing the required computation time, which is particularly important for real-time or near-real-time load forecasting systems. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 Flowchart of the method described in Example 1. ​DETAILED DESCRIPTION

[0039] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1

[0041] This embodiment discloses an ultra-short-term load forecasting method based on CEEMDAN and TSFormer. This method is specifically designed to process and forecast complex power load data, especially ultra-short-term power load forecasting. Ultra-short-term power load forecasting needs to consider the time series characteristics of the data, including short-term and long-term time dependencies. Many machine learning models fail to effectively capture this dependency when processing time series data, resulting in the prediction results not accurately reflecting the load change trend. Although existing technologies have made certain progress in the field of power load forecasting, there are still problems with insufficient prediction accuracy and reliability, especially when predicting load peaks or troughs in the short term, which is a major technical problem that needs to be solved.

[0042] This embodiment proposes a power load forecasting model based on the combination of time series decomposition and sample entropy sequence RNN prediction, which mainly includes three parts: time series decomposition, reconstruction of the decomposed sequence based on sample entropy, and prediction of the reconstructed sequence using a neural network. Figure 1 As shown, the method includes the following steps:

[0043] S01) Time series decomposition: The original load series is decomposed using the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) method. This method effectively decomposes nonlinear and nonstationary power data into multiple IMF component sequences and simultaneously extracts spatiotemporal features, thereby improving the quality and structure of the data and laying the foundation for subsequent feature extraction and load forecasting. The extracted IMF component sequences reveal the different frequency components and trends of the data, which helps to better understand and predict changes in power load. The decomposition results include several IMFs and a residual, each of which corresponds to a specific frequency component in the original signal.

[0044] S02) Reconstruct the IMF component series based on sample entropy. Sample entropy is a method used to measure the complexity of non-stationary time series. Compared to other methods, it is independent of data length and has good consistency. The value of sample entropy is positively correlated with the degree of sequence self-similarity. In the second stage of reconstructing the decomposed series based on sample entropy, we used four different sets of parameters and found that the different parameters had little effect on the trend of the sample entropy results. Reconstructing the previous IMF component series based on sample entropy yielded two new series: the fluctuation series F1 and the trend series F2.

[0045] S03), a short-term load forecasting model based on the TSFormer network is constructed, the fluctuation sequence F1 and the trend sequence F2 are input into the short-term load forecasting model, and ultra-short-term power load forecasting is performed.

[0046] The core of the TSFormer design is to use the self-attention mechanism of the Transformer to capture long-distance dependencies in time series data. Compared with traditional RNN and LSTM, the Transformer can process sequence data in parallel and learn complex patterns in the sequence more efficiently, which is particularly important for predicting short-term and ultra-short-term loads of power systems.

[0047] In this embodiment, step S01) is specifically:

[0048] S11), adding Gaussian noise, adding a limited number of adaptive Gaussian white noise to the original load sequence, and the expression is:

[0049] (1);

[0050] wherein represents the original load sequence, represents the sequence obtained after the original load sequence is added with Gaussian white noise for the i-th time, represents the noise coefficient, represents the Gaussian white noise added for the i-th time;

[0051] S12), obtaining the first IMF component, using the average value of the first IMF component to perform EMD decomposition, as shown in equation (2):

[0052] (2),

[0053] wherein is the first IMF obtained after the original load sequence with Gaussian white noise added for the i-th time is decomposed, M is the number of Gaussian white noise decompositions, and then the original load sequence is subtracted from IMF1 to obtain the first residual sequence, as shown in equation (3):

[0054] (3);

[0055] wherein is the residual sequence obtained;

[0056] S13), obtaining the second IMF, performing EMD decomposition on to obtain the second IMF component, and the result is shown in equation (4):

[0057] (4),

[0058] in 、 They represent the operation of performing the j-th EMD decomposition on the corresponding signal and extracting the j-th IMF components. Represents the weight coefficient after adding Gaussian white noise;

[0059] S14) Calculate the final sequence of residuals and repeat formula (5) to obtain the remaining IMF components.

[0060] (5);

[0061] Where k is the total number of IMF components, 、 represents the kth and k-1th residual sequences, is the kth IMF component, represents the weight coefficient associated with the k-th operation, 、 Respectively represent EMD decomposition of the corresponding signals;

[0062] The final residual sequence is (6), where res is the final residual sequence.

[0063] In the above description, 、 Both represent EMD decomposition of the corresponding signals, but the operation objects are different. Indicates an operation on a specific object or expression, emphasizing the process of applying the operation. Indicates an operation on a set or with specific conditions, emphasizing the integrity or scope of the operation. The superscript j is the number of operations. The specific number j is a step in a loop or iteration process, representing the current jth EMD processing.

[0064] In this embodiment, step S02) is specifically as follows:

[0065] S21), the sample entropy sequence is formed, and the sequence is formed into multiple vectors X with a dimension of m m (1), X m (2),…,X m (N–m+1), (7);

[0066] in represents the i-th IMF component obtained by decomposition, and N represents the number of samples in the time series;

[0067] S22), vector distance calculation, the distance between vectors Xm(i) and Xm(j) can be calculated as follows:

[0068] (8);

[0069] S23) Calculation of Bi: For any Xm(i), we calculate the distance between Xm(i) and Xm(j) to be less than or equal to r. The number of is denoted as Bi, and Bi is defined as follows:

[0070] (9),

[0071] Then the calculation formula of Bi is:

[0072] (10),

[0073] Where r represents the similarity threshold, which is used to determine whether two sequences are similar. When the embedding dimension is m, the number of similar sequences of subsequence x(i) is, represents the average similarity of sequences with embedding dimension m in the entire time series;

[0074] S24) Calculation of Si: After increasing the dimension to m+1, the number of distances between Xm(i) and Xm(j) that are less than or equal to r is expressed as Si, which is defined as follows:

[0075] (11),

[0076] Then the calculation formula of Si is:

[0077] (12),

[0078] in When the embedding dimension is m, the number of similar sequences of subsequence x(i) is expressed as represents the average similarity of the entire time series when the embedding dimension is m;

[0079] S25) Sample entropy calculation: Calculate the sample entropy as follows:

[0080] (13),

[0081] When N is finite, SampEn is estimated as follows:

[0082] (14).

[0083] The short-term load forecasting model based on the TSFormer network described in this embodiment includes an input layer, an encoder layer, and an output layer. The input layer receives the sequence reconstructed by sample entropy. For input, the time series data must first be converted into a format suitable for neural network processing. This typically involves embedding the numerical features of the time series and adding positional encoding to preserve temporal information.

[0084] Feature embedding is the process of converting time series data into an embedded vector in a high-dimensional space. This is usually achieved through a fully connected layer that maps the features of each time point to a vector of fixed size.

[0085] Since the Transformer structure itself does not have the ability to capture position information in the sequence, it is necessary to add position encoding to the embedding vector to provide time order information. The position encoding is learned through a static cosine function.

[0086] The encoder layer is composed of multiple stacked sub-encoder layers, each of which includes a self-attention layer and a feed-forward network.

[0087] The self-attention layer allows the model to consider information from different positions in the sequence when making predictions, enhancing its ability to handle long-range dependencies. Specifically, the position-encoded embedding vector is transformed through different query (Q), key (K), and value (V) matrices before performing a scaled dot-product attention calculation. The self-attention layer uses a multi-head attention mechanism, where the position-encoded embedding vector is transformed in parallel across multiple sets of query, key, and value matrices. The resulting outputs are concatenated and linearly transformed again to produce the final attention output.

[0088] Each self-attention layer is connected to a feed-forward network, which applies the same fully connected layer to the output of each position independently. A common structure is two layers of linear transformation with a ReLU activation function in between.

[0089] The output of each sub-coding layer is added to the input of this layer and then normalized and connected to the residual layer, which helps to avoid the gradient vanishing problem that occurs in deep networks.

[0090] After passing through a series of encoder layers, the output of the short-term load forecasting model typically passes through one or more fully connected layers to make a final prediction, predicting the power load at the next few time points. For power load forecasting, the output layer directly predicts the load at a specific moment in the future.

[0091] In this embodiment, the short-term load forecasting model must be trained first, and the training process is as follows:

[0092] S41): Data division:

[0093] Before training, the dataset needs to be divided into a training set, a validation set, and a test set. The training set is used for model learning, the validation set is used to adjust hyperparameters and prevent model overfitting, and the test set is used to ultimately evaluate the model's performance.

[0094] S42): Batch processing and data loading:

[0095] Data is typically processed in small batches. Using an appropriate batch size can balance training speed and memory usage. Small batches also help models learn more stably. Data loaders (such as PyTorch's DataLoader) can be used to automatically handle data batching, shuffling, and parallel loading.

[0096] S42): Loss calculation:

[0097] The loss is calculated based on the predicted results and the true labels. The loss function is the mean square error (MSE).

[0098] S43): Parameter optimization:

[0099] The Adam optimization algorithm is used to update the model parameters based on the gradient. Optimizers often incorporate momentum and adaptive learning rate techniques to accelerate the training process and improve the convergence rate of the model.

[0100] S5: Verify the accuracy of the prediction model

[0101] Verifying the accuracy of a forecasting model is a critical step in ensuring that the model works reliably in real-world applications. The following are several commonly used methods and metrics to evaluate and verify the accuracy of time series forecasting models.

[0102] S51): Performance evaluation indicators:

[0103] Mean Squared Error (MSE): measures the average of the squares of the differences between the predicted values ​​and the actual values. Root Mean Squared Error (RMSE): the square root of the MSE, providing the error size in the same units as the original data. Mean Absolute Error (MAE): another measure of forecast accuracy, which represents the average of the absolute values ​​of the differences between the predicted values ​​and the actual values. Mean Absolute Percentage Error (MAPE): this is a commonly used proportional indicator used to measure the degree to which the predicted values ​​deviate from the actual values.

[0104] S52): Cross validation:

[0105] Time Series Cross-Validation: For time series data, cross-validation requires appropriate processing to preserve the temporal order. This typically involves a rolling or rolling window approach, where the model is trained on a fixed-size time window and the window is incrementally moved forward.

[0106] S53): Residual analysis:

[0107] Examination of residuals: Analyzing the prediction residuals (i.e., the difference between the predicted and actual values) can reveal whether there are systematic biases or patterns in the model. Ideally, the residuals should appear randomly distributed, with no obvious patterns or trends.

[0108] After ensuring the model's performance meets requirements, deploy it to a production environment to predict power load in real time. Continue to monitor the model's performance and fine-tune it as needed.

[0109] In this embodiment, the input of the ultra-short-term load forecasting model includes, in addition to the original load sequence, power usage, time information (such as date and time), environmental factors (such as temperature and humidity), etc.

[0110] This implementation combines the efficient data decomposition capabilities of CEEMDAN with the powerful time series modeling capabilities of TSFormer to provide a highly accurate and reliable power load forecasting system for the park. This not only enhances the operational efficiency of the power grid but also provides valuable decision support for the park's participation in power market transactions.

[0111] The above description is only the basic principle and preferred embodiments of the present invention. Improvements and substitutions made by those skilled in the art based on the present invention fall within the protection scope of the present invention.

Claims

1. A method for ultra-short-term power load forecasting, characterized by: The following steps are involved: S01) Decompose the original load sequence using the synchronous extension of the fully integrated empirical mode decomposition, decompose multiple IMF component sequences from the nonlinear and non-stationary power data, and extract the spatiotemporal features synchronously; Step S01) is specifically as follows: S11) Add Gaussian noise. Add a limited number of adaptive Gaussian white noises to the original load sequence. The formula is: （1）； in represents the original load sequence, represents the sequence obtained by adding Gaussian white noise to the original load sequence for the i-th time, represents the noise figure, represents the Gaussian white noise added for the i-th time; S12) Find the first IMF component and use the average value of the first IMF component to find the Perform EMD decomposition as shown in formula (2): （2）， in is the first IMF obtained after decomposing the original load sequence after adding Gaussian white noise for the i-th time, M is the number of Gaussian white noise decompositions, and then the original load sequence is used Subtracting IMF1, we get the first residual sequence, as shown in formula (3): （3）； in is the obtained residual sequence; S13) Find the second IMF, Perform EMD decomposition to obtain the second IMF component, and the result is shown in formula (4): （4）， in 、 They represent the operation of performing the j-th EMD decomposition on the corresponding signal and extracting the j-th IMF components. Represents the weight coefficient after adding Gaussian white noise; S14) Calculate the final sequence of residuals and repeat formula (5) to obtain the remaining IMF components. （5）； Where k is the total number of IMF components, 、 represents the kth and k-1th residual sequences, is the kth IMF component, represents the weight coefficient associated with the k-th operation, 、 Respectively represent EMD decomposition of the corresponding signals; The final residual sequence is (6), where res is the final residual sequence; S02), reconstructing the IMF component sequence obtained by decomposition in step S01) based on sample entropy to obtain two new sequences, namely the fluctuation sequence F1 and the trend sequence F2; S03), the fluctuation sequence F1 and the trend sequence F2 are input into the short-term load forecasting model composed of the TSFormer network. The TSFormer network includes an input layer, an encoder layer and an output layer. The input layer performs feature embedding and position encoding on the input sequence. Feature embedding is to convert the input sequence into an embedding vector in a high-dimensional space, and position encoding is to add position encoding to the embedding vector to retain time information; the encoder layer is composed of multiple sub-encoding layers stacked together, each of which includes a self-attention layer and a feedforward network. The self-attention layer converts the position-encoded embedding vector through different query, key and value matrices, and then performs scaled dot product attention calculation to obtain the final attention output. Each self-attention layer is followed by a feedforward network, and the feedforward network independently applies the same fully connected layer to the output of each position for linear transformation and activation; the output layer is located after the encoder layer, and performs the final prediction through one or more fully connected layers to predict the power load values ​​at the next few time points or a certain moment.

2. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The self-attention layer adopts a multi-head self-attention mechanism, that is, the position-encoded embedding vector is transformed in parallel on multiple sets of different query, key, and value matrices, and the output results are spliced ​​and linearly transformed again to obtain the final attention output.

3. The ultra-short-term power load forecasting method according to claim 1, characterized in that: Feature embedding is achieved through a fully connected layer, which maps the features of each time point to a vector of fixed size.

4. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The positional encoding is learned via a static cosine function.

5. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The output of each sub-coding layer is added to the input of the layer and then normalized and connected with the residual.

6. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The feedforward network uses two layers of linear transformation with a ReLU activation function in the middle.

7. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The input of the short-term load forecasting model also includes weather data and date type.

8. The ultra-short-term power load forecasting method according to claim 1, characterized in that: The decomposition result of step S01) includes multiple IMF component sequences and a residual, and each IMF component sequence corresponds to a specific frequency component in the original load signal.

9. An ultra-short-term power load forecasting device, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the short-term power load forecasting method according to any one of claims 1 to 8 when running the program instructions.

Images