Short-term power load prediction method based on time sequence periodic feature self-learning
By employing a time-series periodic feature self-learning method based on variational inference and mutual information constraints, the power load sequence was successfully decomposed into periodic and trend components. Combined with the CNN-LSTM model, the accuracy and robustness issues of power load forecasting were resolved, thus improving the forecasting performance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN UNIV OF SCI & TECH
- Filing Date
- 2026-01-21
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to effectively decompose electricity load time series into clear periodic and trend components, limiting the accuracy and robustness of AI-based prediction models.
A self-learning method based on time series periodic features is adopted. Through variational inference framework and mutual information constraints, the power load time series is automatically learned and decomposed into periodic and trend latent variables, and then combined with convolutional neural network and long short-term memory network for prediction.
It significantly improves the accuracy and robustness of short-term power load forecasting, enhances the model's resistance to noise and abnormal fluctuations, simplifies the data processing flow, and improves its applicability in complex real-world scenarios.
Smart Images
Figure CN121965501A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load forecasting technology, and specifically to a short-term power load forecasting method based on self-learning of time series periodic characteristics. Background Technology
[0002] With the deepening of electricity substitution and the continuous improvement of electrification levels, load characteristics in the new power system architecture have evolved into a multi-dimensional coupled dynamic system. Accompanying the market-oriented transformation of electricity trading mechanisms, innovative business models such as demand-side response mechanisms, resource-sharing paradigms, load aggregation operation entities, and distributed energy clusters are constantly emerging, causing load composition to exhibit time-varying and heterogeneous characteristics. In recent years, due to the widespread deployment of smart metering devices and the rapid improvement of communication capabilities, power grid companies have accumulated a large amount of multi-source heterogeneous data through various information platforms, providing an important data foundation for the application of cutting-edge technologies such as artificial intelligence (AI) and big data analysis in the field of power load forecasting. However, power load time series are usually composed of multiple intertwined and complex components, making the process of decomposing them into clear and static load sequences more complex, fundamentally limiting the effectiveness of AI-based forecasting models.
[0003] In summary, there is an urgent need in the existing technology for a method to learn the inherent periodic trend characteristics of power system load time series based on artificial intelligence technology in order to solve this problem. Summary of the Invention
[0004] The purpose of this invention is to overcome the problems in the prior art and provide a short-term power load forecasting method based on self-learning of time series periodic features. This method can effectively decompose complex power load time series into clear periodic and trend components, thereby improving the accuracy and robustness of the forecast results.
[0005] This invention provides a short-term power load forecasting method based on self-learning of time series periodic features, comprising: Acquire historical time-series data of electricity load, along with associated meteorological and date-type data; The historical time series data is input into a periodic trend feature learning model based on variational inference. This model automatically learns and decomposes a first latent variable to characterize the inherent periodic characteristics of the time series data and a second latent variable to characterize its trend characteristics by optimizing the lower bound of evidence. The first hidden variable, the second hidden variable, the meteorological data, and the date-type data are input into the hybrid prediction model, and the predicted value of the power load for the future period is output. The objective function of the periodic trend feature learning model includes a mutual information constraint, which is configured to maximize the mutual information between the first latent variable, the second latent variable and the original time series data, while minimizing the mutual information between the first latent variable and the second latent variable, so that the first latent variable and the second latent variable are independent of each other.
[0006] Preferably, the lower bound objective of the evidence includes an expected reconstruction loss and a KL divergence regularization term; the expected reconstruction loss is used to make the data reconstructed from the latent variables as close as possible to the original time series data; the KL divergence regularization term is used to make the approximate posterior distribution of the latent variables as close as possible to a preset prior distribution; the lower bound objective of the evidence is optimized by maximizing the expected reconstruction loss and minimizing the KL divergence regularization term.
[0007] Preferably, the expected reconstruction loss is decomposed into a first reconstruction loss term and a second reconstruction loss term; the first reconstruction loss term corresponds to the error between the periodic component reconstructed from the first latent variable and the original sequence; the second reconstruction loss term corresponds to the error between the trend component reconstructed from the second latent variable and the original sequence.
[0008] Preferably, the calculation of the first reconstruction loss term and the second reconstruction loss term is based on the Gaussian distribution assumption and is achieved by combining the autocorrelation coefficient and time correlation coefficient of the original time series data.
[0009] Preferably, the hybrid prediction model includes a convolutional neural network module and a long short-term memory network module; the convolutional neural network module is used to extract features from the meteorological data and date-type data organized by time series, and output an external feature sequence with the same number of time steps as the historical load sequence; the long short-term memory network module is used to receive the fused feature sequence and perform time series modeling, wherein the fused feature sequence is formed by concatenating the first latent variable, the second latent variable and the corresponding features in the external feature sequence at the same time step, and finally outputs the load prediction result for the future period.
[0010] Preferably, the meteorological data includes one or more of temperature, humidity, wind speed, and weather type.
[0011] The present invention also discloses the above-mentioned short-term power load forecasting system, comprising: The data acquisition module is used to acquire historical time series data of power load, as well as meteorological data and date-type data associated with the time series; A periodic trend feature learning module is used to receive the historical time series data and process the data based on a variational inference framework. The module automatically learns and decomposes a first latent variable and a second latent variable from the historical time series data by optimizing an evidence lower bound objective. The first latent variable characterizes the inherent periodicity of the time series, and the second latent variable characterizes its trend characteristics. The evidence lower bound objective includes a mutual information constraint, configured to maximize the mutual information between the first and second latent variables and the original time series data during training, while minimizing the mutual information between the first and second latent variables, so that they are independent of each other. The hybrid forecasting module is used to receive the first latent variable, the second latent variable, the meteorological data, and the date-type data. The hybrid forecasting module includes a feature extraction unit and a time-series forecasting unit. The feature extraction unit is used to extract features from the meteorological data and the date-type data to obtain an external feature vector. The time-series forecasting unit is used to receive the external feature vector, the first latent variable, and the second latent variable, and perform time-series modeling to output the predicted power load value for future periods.
[0012] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the short-term power load forecasting method based on time series periodic features self-learning as described above.
[0013] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the short-term power load forecasting method based on time series periodic features self-learning as described in any one of the above-mentioned methods.
[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a short-term power load forecasting method based on variational inference. The periodic trend feature learning model in this method optimizes and integrates the evidence lower bound objective of the decomposition strategy. The model can adaptively learn and decouple the first and second latent variables, representing periodic and trend characteristics respectively, from complex historical load sequences, replacing the original mixed sequences as prediction inputs and laying the foundation for improved accuracy. Secondly, this invention also introduces mutual information constraints, maximizing the retention of original information by each latent variable while forcing them to be independent, thereby enhancing the model's robustness to noise and abnormal fluctuations. This invention utilizes the probabilistic modeling framework of variational autoencoders, mapping the sequence to a regular latent variable space and achieving automatic decoupling through approximate posterior distribution and KL divergence regularization. This fundamentally avoids the manual intervention problem of traditional deterministic decomposition methods and simplifies the data processing flow. Finally, this method significantly outperforms traditional AI models and their combinations with fixed decomposition algorithms in both prediction accuracy and model robustness, greatly improving its applicability in complex real-world power scenarios. Attached Figure Description
[0015] Figure 1 This is a flowchart of the LiPTR prediction method based on self-learning of time series periodic features according to the present invention.
[0016] Figure 2 This is a structural diagram of the CNN-LSTM hybrid prediction model of the prediction method based on time series periodic features of this invention.
[0017] Figure 3 This is a graph of power load data from the prediction method based on time series periodic features of this invention.
[0018] Figure 4 This is a comparison chart of the prediction results of the prediction method based on self-learning of time series periodic features in this invention.
[0019] Figure 5 This is a reconstruction diagram of the periodic trend component of the prediction method based on self-learning of time series periodic features according to the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0021] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as “comprising” or “including” indicate that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; when the absolute position of the described objects changes, the relative positional relationship may also change accordingly.
[0022] This embodiment provides a short-term power load forecasting method based on self-learning of time series periodic features, including: acquiring historical time series data of power load as well as associated meteorological data and date-type data; The historical time series data is input into a periodic trend feature learning model based on variational inference. This model automatically learns and decomposes a first latent variable to characterize the inherent periodic characteristics of the time series and a second latent variable to characterize its trend characteristics by optimizing an evidence lower bound objective. The first hidden variable, the second hidden variable, the meteorological data, and the date-type data are input into the hybrid prediction model, and the predicted value of the power load for the future period is output. The objective function of the periodic trend feature learning model includes a mutual information constraint. This constraint is configured to maximize the mutual information between the first latent variable, the second latent variable, and the original time series data, while minimizing the mutual information between the first latent variable and the second latent variable, to ensure that the first latent variable and the second latent variable are independent of each other. Preferably, the weight of the constraint term minimizing the mutual information between latent variables should be greater than the weight of the constraint term maximizing the mutual information between latent variables and the data, as well as the weight of the KL divergence term.
[0023] The periodic trend feature learning model of this invention is based on a variational inference framework, training a parameterized model by optimizing the Evidence Lower Bound (ELBO) objective. This model takes historical time series data as input, maps it to a latent variable space through an encoding network, and automatically decouples the latent variables into two independent semantic components during training: a first latent variable and a second latent variable. The first latent variable corresponds to the periodic component in this embodiment, and the second latent variable corresponds to the trend component. In this decoupling process, information theory-based mutual information constraints are introduced into the objective function. This ensures that the first and second latent variables retain the periodic and trend information from the original sequence, maximizing their mutual information with the original data, while simultaneously minimizing their mutual information, achieving a clear and adaptive separation of periodic and trend features. The two separated latent variable features are then fused with external influencing factors such as weather and date type extracted by a convolutional neural network, and finally input into a long short-term memory network for joint time series modeling to complete load forecasting.
[0024] In this embodiment of the invention, because the feature learning process is directly coupled with the final prediction task, the extracted periodic and trend components can more effectively support the prediction model. Compared to traditional methods... EMD, EWT Decomposition and LSTM Compared with other methods, the prediction accuracy of the embodiments of the present invention is superior.
[0025] The prediction model of this invention can adaptively learn the inherent patterns of load data with different characteristics in different regions. It has better adaptability to complex situations such as non-stationarity of load sequences, holiday effects, and abnormal weather, and improves stability and generalization in variable real-world scenarios.
[0026] This invention utilizes mutual information constraints and variational reasoning to achieve semantic and decoupled representation learning of temporal components, avoiding problems such as modality mixing and boundary effects that may occur in traditional decomposition methods, and providing a certain interpretability basis for model prediction results.
[0027] As another preferred embodiment, the lower bound of evidence in this embodiment includes the expected reconstruction loss and KL The divergence regularization term; the expected reconstruction loss is used to make the data reconstructed from the latent variables as close as possible to the original time series data; the KL The divergence regularization term is used to make the approximate posterior distribution of the latent variable as close as possible to a preset prior distribution; the evidence lower bound objective is achieved by maximizing the expected reconstruction loss and minimizing the... KL Optimization is achieved using divergence regularization terms.
[0028] In this embodiment, the expected reconstruction loss measures the ability to decode the learned latent variables back to the original time series data. Its purpose is to force the latent variables learned by the model to contain enough information to accurately recover the original sequence, thereby ensuring that the latent variables can effectively represent the core features of the input data. KL The divergence regularization term constrains the approximate posterior distribution of latent variables to approach a pre-defined simple prior distribution. Its function is to normalize the structure of the latent variable space, prevent the model from overfitting noise or irrelevant details in the training data, and improve the regularity and generalization ability of the learned features. This embodiment enhances representational ability by maximizing the expected reconstruction loss while minimizing... KL The divergence regularization term maintains the normality of the distribution, such as satisfying the standard normal distribution. The model can achieve an optimal balance between restoring the data as much as possible and keeping the latent variable structure simple and orderly.
[0029] In another preferred embodiment, the desired reconstruction loss is decomposed into a first reconstruction loss term and a second reconstruction loss term; the first reconstruction loss term corresponds to the error between the periodic component reconstructed from the first latent variable and the original sequence; the second reconstruction loss term corresponds to the error between the trend component reconstructed from the second latent variable and the original sequence.
[0030] This invention, through task-oriented decomposition of the reconstruction loss, enables the model to more accurately and efficiently separate the mixed periodic and trend information in the load sequence, improving the clarity and purity of feature decoupling. Simultaneously, in this embodiment, the periodic and trend latent variables converge stably on their own, facilitating the independent utilization of these two types of features by the subsequent prediction model. This also helps improve the accuracy and robustness of the final load prediction.
[0031] In another preferred embodiment, the calculation of the first reconstruction loss term and the second reconstruction loss term is based on the Gaussian distribution assumption and is achieved by combining the autocorrelation coefficient and time correlation coefficient of the original time series data. This embodiment binds abstract periodic and trend semantics with computable statistical properties, such as autocorrelation and time correlation, providing the model with a clear and differentiable optimization direction. This makes the learning process more stable and efficient, and enables the decoupled feature components to possess the desired statistical properties and physical interpretability.
[0032] In another preferred embodiment, the hybrid prediction model includes a convolutional neural network module and a long short-term memory network module. The convolutional neural network module is used to extract features from the meteorological data and date-type data organized by time series, and output an external feature sequence with the same number of time steps as the historical load sequence. The long short-term memory network module is used to receive the fused feature sequence and perform time series modeling. The fused feature sequence is formed by concatenating the first latent variable, the second latent variable, and the corresponding features in the external feature sequence at the same time step, and finally outputs the load prediction result for the future period. This embodiment uses... CNN and LSTM The clear division of labor enables optimal feature extraction and fusion of the load's inherent patterns and external factors, overcoming the inefficiency or information loss issues associated with processing multi-source data using a single network structure. Furthermore, this embodiment uses the decoupled periodic and trend latent variables as direct inputs, providing... LSTM It provides clear and pure temporal semantic information, reducing the difficulty for the model to learn long-term dependencies from the original mixed data, and improving modeling efficiency and prediction accuracy.
[0033] In a preferred embodiment, the meteorological data includes one or more of temperature, humidity, wind speed, and weather type.
[0034] This invention also provides the above-mentioned short-term power load forecasting system, which includes: The data acquisition module is used to acquire historical time series data of power load, as well as meteorological data and date-type data associated with the time series; A periodic trend feature learning module is used to receive the historical time series data and process the data based on a variational inference framework. The module automatically learns and decomposes a first latent variable and a second latent variable from the historical time series data by optimizing an evidence lower bound objective. The first latent variable characterizes the inherent periodicity of the time series, and the second latent variable characterizes its trend characteristics. The evidence lower bound objective includes a mutual information constraint, configured to maximize the mutual information between the first and second latent variables and the original time series data during training, while minimizing the mutual information between the first and second latent variables, so that they are independent of each other. The hybrid forecasting module is used to receive the first latent variable, the second latent variable, the meteorological data, and the date-type data. The hybrid forecasting module includes a feature extraction unit and a time-series forecasting unit. The feature extraction unit is used to extract features from the meteorological data and the date-type data to obtain an external feature vector. The time-series forecasting unit is used to receive the external feature vector, the first latent variable, and the second latent variable, and perform time-series modeling to output the predicted power load value for future periods.
[0035] More specifically, the prediction method based on self-learning of time series periodic features in this embodiment includes the following steps, which are performed sequentially. Step 1: For time series datasets X 1:T = { X 1, X 2, …, X t , …, X T}, LiPTR The goal is to obtain a feature that can be used to predict future time series. Z and through features Z Reconstructing the cycle-trend time series components: LiPTR Hoping to use historical time series X Obtain features Z And obtain future time series through prediction Y Maximizing the probability, i.e., maximizing the log-likelihood log P(X,Y) : ; Since the true posterior distribution cannot be directly obtained P(Z|X) ,feature Z The distribution of is still unclear, which leads to log P (X,Y) The calculation becomes extremely complex. To address this issue, this embodiment starts from variational inference theory and introduces an approximate posterior distribution. Q φ ( Z | X To replace the true posterior distribution P(Z|X) This transforms the original posterior inference problem into an optimization problem, and through... Jensen The inequality leads to a lower bound on the log-likelihood:
[0036] ; This can be further simplified to the form of expectation, through which it becomes clearer how to maximize the lower bound of the log-likelihood: ; KL Divergence measures the difference between two probability distributions, and its value is non-negative. It is achieved by minimizing... KL Divergence can approximate the posterior distribution. Q φ ( Z | X As close as possible to the prior distribution P ( Z This ensures that the obtained features Z It follows a standard normal distribution:
[0037] Therefore, the lower bound of evidence is obtained ( ELBO The expression for ) is: ; Where Θ is , and Composition, respectively representing LiPTR The prediction part, the reconstruction part and KL The parameters to be optimized in divergence.
[0038] Step Two, in LiPTR By incorporating a decomposition strategy, features that can be used for time series prediction can be included. Z Correspondingly decomposed into Z P and Z T and obtained LiPTR Objective function: By adopting a decomposition strategy ELBO It has the following decomposition forms: Under the Gaussian distribution assumption, there is no need to utilize X P and X T The reconstruction loss can then be calculated. ELBO The reconstructed portion can be replaced by the following formula: ; in, It is an autocorrelation coefficient with a lag value κ, which is selected based on the resolution of the power load time series data. For example, for data sampled at 15-minute intervals, κ=96 can be selected to correspond to its daily cycle. It is the time correlation coefficient, where .
[0039] Step 3, to ensure ZP and Z T They are independent of each other, and additional mutual information terms are introduced for constraint: By increasing Z P , Z T and X Mutual information between them can alleviate the problem of feature divergence reduction, while reducing Z P and Z T Mutual information between them, to further separate their features, to ensure Z P and Z T They are independent of each other. Therefore, LiPTR The objective function is:
[0040] ; In the formula It represents the mutual information between two variables.
[0041] Step 4: Construction CNN-LSTM Mixed prediction models; Weather factors have a crucial impact on short-term load forecasting. Among these factors, the most common is temperature, followed by humidity, wind speed, and weather type. Summer and cold winters can lead to drastic load variations. Date type is another important factor influencing short-term forecasting. Current electricity load is primarily driven by industrial consumption, with significantly reduced load on non-working days (such as holidays, Saturdays, and Sundays) compared to weekdays (Monday to Friday). In embodiments of the invention, characteristic quantities from load time series data are used. Z P and Z T Meteorological data and date type factors are used as inputs for model training.
[0042] The hybrid predictor consists of two main parts: CNN and LSTM The former is primarily responsible for extracting influencing factor features from weather data and date types, while the latter is primarily responsible for load forecasting. Its inputs include... CNN The output and the feature values obtained after encoding the load time series data. Z P and Z T . Z P and Z TAll are feature sequences with the same number of time steps as the input historical load sequence. The CNN module extracts features from meteorological data and date-type data, and the output is the same as... Z P and Z T External feature sequences with the same number of time steps. Before inputting into the LSTM module, the external feature sequences with the same time step... Z P and Z T The fusion feature sequence is concatenated with the external feature vector to form a fused feature sequence, which is used as the input of the LSTM module for time series modeling, and finally outputs the load forecast value for future periods.
[0043] This invention proposes a novel short-term power load forecasting method based on learning the inherent periodic trend characteristics of load sequences using a variational autoencoder. The first step involves using a variational autoencoder to decompose the load sequence into different periodic and trend components, thereby generating separately predictable periodic and trend components. Secondly, a specialized hybrid forecasting model is introduced. CNN-LSTM This is used to predict cyclical and trend components. Normalized date type and weather characteristics are also used as additional inputs.
[0044] More specifically, the periodic trend feature learning model in this embodiment adopts a variational autoencoder architecture, which mainly consists of an encoder and a decoder.
[0045] In this embodiment, the encoder consists of a bidirectional long short-term memory network. Its input is a standardized historical load time series, and the hidden state of the last time step is mapped via a fully connected layer, outputting two sets of parameters: a first hidden variable used to characterize the periodicity. Z P The mean and log-variance, and a second latent variable used to characterize the trend. Z T The mean and log-variance of the variables are obtained by sampling from their respective Gaussian distributions using reparameterization techniques. Z P and Z T .
[0046] The decoder is a multilayer perceptron. Its input is the sampled latent variables. Z P and Z T The concatenated vector is passed through two fully connected layers (with ReLU activation function in the middle layer), and the final output layer reconstructs the data into a reconstructed sequence with the same dimensions as the original sequence. The reconstructed sequence can be viewed as the sum of the periodic component and the trend component. In another implementation, the decoder can also be designed as a dual-path structure, independently reconstructing the periodic component and the trend component.
[0047] To achieve the aforementioned mutual information constraint, this embodiment employs a neural network-based mutual information estimation and optimization technique during training. Specifically, to maximize the mutual information between latent variables and the data, an auxiliary discriminant network is introduced, enabling it to distinguish between true samples from the joint distribution and negative samples constructed from the product of marginal distributions. The encoder is trained to improve the error rate of the discriminant network, thereby increasing the lower bound of the neural estimation of mutual information.
[0048] Minimize the mutual information between periodic and trend latent variables: In a similar manner, train the encoder to make it difficult for another discriminative network to distinguish between joint samples and randomly paired marginal samples, thereby reducing its mutual information estimate.
[0049] The above process can be achieved using well-known deep learning paradigms such as maximizing the Jensen-Shannon divergence or optimizing the information-noise contrastive estimation loss. The discriminator network can be simplified to a multilayer perceptron. The strength of the two constraints is controlled by hyperparameters α and γ, respectively; to achieve effective decoupling, γ > α is usually required.
[0050] The present invention will be further described in detail below with reference to embodiments: Two real-world power load data curves are as follows Figure 3 As shown. To study the impact of different decomposition algorithms on power load forecasting, this invention compares... EMD, EWT and LiPTR Performance in power load forecasting. (Through...) EMD and EWT The subsequences obtained by decomposition and through LiPTR The periodic trend representation obtained from the decomposition was used as model input. Then, the predictive performance of the three decomposition algorithms was compared. Table 1 lists the results. EMD, EWT and LiPTR The short-term power load forecasting accuracy of three decomposition algorithms was evaluated. Furthermore, a representative time window was selected to visually display the forecast results, such as... Figure 4 As shown.
[0051] The invention was further validated by visually reconstructing the periodic and trend components in the data. LiPTR .like Figure 5 As shown, LiPTR It can effectively capture periodic patterns on real-world datasets. Furthermore, trend and periodic components can accurately reconstruct the initial sequence from their respective perspectives, demonstrating that this invention can generate effective features for complex time series.
[0052] Table 1 EMD, EWT and LiPTR Comparison of prediction accuracy Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A short-term power load forecasting method based on self-learning of time series periodic features, characterized in that, include: Acquire historical time-series data of electricity load, along with associated meteorological and date-type data; The historical time series data is input into a periodic trend feature learning model based on variational inference. This model automatically learns and decomposes a first latent variable to characterize the inherent periodic characteristics of the time series data and a second latent variable to characterize its trend characteristics by optimizing the lower bound of evidence. The first hidden variable, the second hidden variable, the meteorological data, and the date-type data are input into the hybrid prediction model, and the predicted value of the power load for the future period is output. The objective function of the periodic trend feature learning model includes a mutual information constraint, which is configured to maximize the mutual information between the first latent variable, the second latent variable and the original time series data, while minimizing the mutual information between the first latent variable and the second latent variable, so that the first latent variable and the second latent variable are independent of each other.
2. The short-term power load forecasting method based on self-learning of time series periodic features as described in claim 1, characterized in that, The lower bound objective of the evidence includes the expected reconstruction loss and the KL divergence regularization term; the expected reconstruction loss is used to make the data reconstructed from the latent variables as close as possible to the original time series data; the KL divergence regularization term is used to make the approximate posterior distribution of the latent variables as close as possible to a preset prior distribution; the lower bound objective of the evidence is optimized by maximizing the expected reconstruction loss and minimizing the KL divergence regularization term.
3. The short-term power load forecasting method based on self-learning of time series periodic features as described in claim 2, characterized in that, The expected reconstruction loss is decomposed into a first reconstruction loss term and a second reconstruction loss term; the first reconstruction loss term corresponds to the error between the periodic component reconstructed from the first latent variable and the original sequence; the second reconstruction loss term corresponds to the error between the trend component reconstructed from the second latent variable and the original sequence.
4. The short-term power load forecasting method based on self-learning of time series periodic features as described in claim 3, characterized in that, The calculation of the first reconstruction loss term and the second reconstruction loss term is based on the Gaussian distribution assumption and is achieved by combining the autocorrelation coefficient and time correlation coefficient of the original time series data.
5. The short-term power load forecasting method based on self-learning of time series periodic features as described in claim 1, characterized in that, The hybrid prediction model includes a convolutional neural network module and a long short-term memory network module. The convolutional neural network module is used to extract features from the meteorological data and date-type data organized by time series, and output an external feature sequence with the same number of time steps as the historical load sequence. The long short-term memory network module is used to receive the fused feature sequence and perform time series modeling. The fused feature sequence is formed by concatenating the first latent variable, the second latent variable and the corresponding features in the external feature sequence at the same time step, and finally outputs the load prediction result for the future period.
6. The short-term power load forecasting method based on self-learning of time series periodic features as described in claim 1, characterized in that, The meteorological data includes one or more of the following: temperature, humidity, wind speed, and weather type.
7. A short-term power load forecasting system, characterized in that, include: The data acquisition module is used to acquire historical time series data of power load, as well as meteorological data and date-type data associated with the time series; A periodic trend feature learning module is used to receive the historical time series data and process the data based on a variational inference framework. The module automatically learns and decomposes a first latent variable and a second latent variable from the historical time series data by optimizing an evidence lower bound objective. The first latent variable characterizes the inherent periodicity of the time series, and the second latent variable characterizes its trend characteristics. The evidence lower bound objective includes a mutual information constraint, configured to maximize the mutual information between the first and second latent variables and the original time series data during training, while minimizing the mutual information between the first and second latent variables, so that they are independent of each other. The hybrid forecasting module is used to receive the first latent variable, the second latent variable, the meteorological data, and the date-type data. The hybrid forecasting module includes a feature extraction unit and a time-series forecasting unit. The feature extraction unit is used to extract features from the meteorological data and the date-type data to obtain an external feature vector. The time-series forecasting unit is used to receive the external feature vector, the first latent variable, and the second latent variable, and perform time-series modeling to output the predicted power load value for future periods.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the short-term power load forecasting method based on self-learning of time series periodic features as described in any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the short-term power load forecasting method based on self-learning of time series periodic features as described in any one of claims 1 to 7.