A deep generative comprehensive energy system source-load prediction method and system

Abstract

The application relates to the technical field of comprehensive energy systems, and provides a deep generative comprehensive energy system source-load prediction method and system, which comprises the following steps: acquiring source-load data and meteorological data of a comprehensive energy system at a plurality of continuous historical time points, and obtaining source-load data at a future time point through a main prediction model; wherein the training step of the main prediction model comprises the following steps: taking real samples in a training set as a basis, taking a discriminator as an opponent of a generator, and guiding the generator to learn to generate virtual samples through adversarial training; after the real samples and the virtual samples are combined, the combined samples are input into the main prediction model, the main prediction model outputs an input discriminator, the authenticity of a prediction result is judged by using the discriminator, and the main prediction model training is guided. The generalization ability of a target model can be significantly improved, and the prediction accuracy of source-load prediction can be improved.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system technology, and in particular relates to a deep-generative integrated energy system source-load prediction method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Energy development must balance both efficient utilization and environmental protection. As a typical paradigm of energy systems, integrated energy systems (IES) can achieve multi-energy complementarity and integrated generation, grid, load, and storage based on advanced physical information technology, thus possessing significant advantages in improving energy efficiency and promoting renewable energy development. Accurate source-load forecasting is the primary prerequisite for IES planning, operation control, and scheduling. However, with the gradual increase in renewable energy penetration and the rapid popularization of loads such as electric vehicles, the source-load characteristics of IES have become more diverse and dynamic, posing significant challenges to forecasting.

[0004] Thanks to the collaborative efforts of numerous scholars, IES (Enhanced Energy Systems) prediction has gradually moved beyond the limitations of directly applying power system prediction methods. It has evolved from early statistical learning, through machine learning, and ultimately to a new stage dominated by deep learning. In recent years, deep learning-based prediction models have primarily used sequence-based neural networks as their backbone to capture time-dependent features, supplemented by convolutional neural networks, attention mechanisms, and multi-task learning (MTL) frameworks to uncover complex coupling relationships between multiple variables. Representative components include Convolutional Neural Networks (CNNs), Residual Networks (ResNets), Transformer networks, Long Short-Term Memory (LSTM) networks, and temporal and spatial attention mechanisms. Among these, the Transformer, based on a self-attention mechanism, exhibits higher parallel computing power and stronger feature representation capabilities when processing long-sequence data, and has rapidly become the mainstream method in the prediction field since its inception. However, wind and solar renewable energy sources and multi-energy loads possess strong randomness and periodicity. Under the IES multi-energy complementary mechanism, these two are coupled with each other, forming a complex time-series structure with multiple layers and dynamic relationships. The Transformer primarily focuses on long-term correlations within sequences, neglecting the coupling relationships between different sequences and the description of local features. Overall, supported by various backbone models, researchers have achieved accurate prediction results through reasonable architecture design. However, most studies still focus on improving the network architecture itself, with less in-depth discussion of training mechanisms and data modeling.

[0005] Data is the primary prerequisite for driving model learning, and its quality and scale directly determine the upper limit of model performance. Especially in the field of IES (Environmentally, Energy Systems), complex dynamic coupling relationships exist between source-load and multi-energy loads, exhibiting high time-varying and multivariate coupling. Deep learning models can only effectively learn their complex mapping characteristics with the support of sufficient high-quality historical data. Besides data constraints, the setting of the optimization objective is also a significant bottleneck limiting the performance of prediction models. Most existing IES prediction methods rely on only a fixed regression optimization objective, such as mean squared error, L1 loss, L2 loss, and Huber loss (smoothed L1 loss). This inflexible optimization objective easily leads to the model overfitting local patterns in historical trajectories, neglecting the modeling of the overall sequence distribution, and making it difficult to adapt to the strong randomness of source-load data. Summary of the Invention

[0006] To address the technical problems existing in the background art, the present invention provides a deep generative integrated energy system source-load prediction method and system, and proposes a deep generative prediction framework, which consists of an auxiliary trainer and a master prediction model. Through the training process, the generalization ability of the target model and the prediction accuracy of source-load prediction can be significantly improved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The first aspect of this invention provides a deep generative integrated energy system source-load prediction method, comprising:

[0009] Acquire source-load data and meteorological data of the integrated energy system at several consecutive historical moments;

[0010] Based on source-load data and meteorological data from several consecutive historical moments, source-load data for future moments are obtained through the master prediction model.

[0011] The training steps of the master prediction model include: using real samples in the training set as a basis, the discriminator acts as the adversarial object of the generator, and the generator learns to generate virtual samples through adversarial training; after merging real samples and virtual samples, the data is input into the master prediction model, and the output of the master prediction model is input into the discriminator. The discriminator is used to judge the authenticity of the prediction results and guide the training of the master prediction model.

[0012] Furthermore, both the generator and the main prediction model employ an encoder. The encoder uses an embedding layer to map the input to a fixed dimension, and then processes it in parallel through two branches. The output of the embedding layer and the outputs of the two branches are added together and then passed sequentially through layer normalization, learnable position encoding, multi-head attention layer, and pre-feedback layer to obtain the encoder output. Residual connections and layer normalization are then used after the multi-head attention layer and the pre-feedback layer.

[0013] Furthermore, the training of the main prediction model employs a weighted sum of weighted multi-task loss and adversarial regularization loss, and the weights of the adversarial regularization loss are: ;in, , and These are the maximum adversarial loss weight, the minimum adversarial loss weight, and the baseline adversarial loss weight, respectively. For gradient direction similarity, , To weighted multi-task loss, To counteract regularization loss.

[0014] Furthermore, the source-load data includes electrical load, thermal load, photovoltaic power generation, and wind power generation.

[0015] Furthermore, the meteorological data includes average temperature, wind speed, and irradiance.

[0016] A second aspect of the present invention provides a deep-generative integrated energy system source-load prediction system, comprising:

[0017] The training module is configured as follows: based on real samples in the training set, the discriminator acts as the adversarial object of the generator, and the generator learns to generate virtual samples through adversarial training; after merging real samples and virtual samples, the results are input into the main prediction model, and the output of the main prediction model is input into the discriminator. The discriminator is used to judge the authenticity of the prediction results and guide the training of the main prediction model.

[0018] The data acquisition module is configured to acquire source-load data and meteorological data of the integrated energy system at several consecutive historical moments.

[0019] The prediction module is configured to obtain future source-load data based on source-load data and meteorological data from several consecutive historical moments, using the master prediction model.

[0020] Furthermore, both the generator and the main prediction model employ an encoder. The encoder uses an embedding layer to map the input to a fixed dimension, and then processes it in parallel through two branches. The output of the embedding layer and the outputs of the two branches are added together and then passed sequentially through layer normalization, learnable position encoding, multi-head attention layer, and pre-feedback layer to obtain the encoder output. Residual connections and layer normalization are then used after the multi-head attention layer and the pre-feedback layer.

[0021] Furthermore, the training of the main prediction model employs a weighted sum of weighted multi-task loss and adversarial regularization loss, and the weights of the adversarial regularization loss are: ;in, , and These are the maximum adversarial loss weight, the minimum adversarial loss weight, and the baseline adversarial loss weight, respectively. For gradient direction similarity, , To weighted multi-task loss, To counteract regularization loss.

[0022] Furthermore, the source-load data includes electrical load, thermal load, photovoltaic power generation, and wind power generation.

[0023] Furthermore, the meteorological data includes average temperature, wind speed, and irradiance.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This invention proposes a deep generative prediction framework consisting of an auxiliary trainer and a master prediction model. Through reinforcement during the training process, it can significantly improve the generalization ability of the target model and the prediction accuracy of source-load prediction.

[0026] Based on the main regression loss function, this embodiment of the invention constructs an adversarial regularization loss based on a discriminator, aiming to model the global structure and deeper levels of the input data. Through the adversarial regularization loss, the model can not only fit the training data more accurately, but also improve its adaptability to unknown data. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a flowchart of the deep generative integrated energy system source-load prediction method according to Embodiment 1 of the present invention;

[0029] Figure 2 This is a structural diagram of the encoder according to Embodiment 1 of the present invention;

[0030] Figure 3 This is a comparison chart of the electrical load prediction results of Embodiment 1 of the present invention;

[0031] Figure 4 This is a comparison chart of the heat load prediction results of Embodiment 1 of the present invention;

[0032] Figure 5 This is a comparison chart of photovoltaic power generation prediction results in Embodiment 1 of the present invention;

[0033] Figure 6 This is a comparison chart of wind power generation prediction results according to Embodiment 1 of the present invention. Detailed Implementation

[0034] 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.

[0035] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0036] Example 1

[0037] This embodiment provides a deep generative integrated energy system source-load prediction method.

[0038] This embodiment provides a deep generative integrated energy system source-load prediction method, which solves the problems of data scarcity and localization of optimization objectives faced by current deep learning prediction models in the training phase through a deep generative prediction framework.

[0039] This embodiment provides a deep generative integrated energy system source-load prediction method. Inspired by the success of generative adversarial networks (GANs) in terms of generation and regularization, it addresses the shortcomings of data scarcity and optimization target localization in the prediction model training process by proposing a deep generative prediction framework (DGP). This framework consists of an auxiliary trainer and a main prediction model, which can be flexibly integrated into any existing prediction model. Through reinforcement during the training process, it can significantly improve the prediction accuracy and generalization ability of the target model.

[0040] Generative Adversarial Networks (GANs) are generative models based on game theory, consisting of a generator and a discriminator. They enhance the model's ability to characterize data distribution information through adversarial optimization, thereby generating high-fidelity samples consistent with real data. Inspired by this, this embodiment introduces a generative adversarial mechanism into a time series prediction task, constructing an auxiliary trainer to enhance the training process from two perspectives: data augmentation and optimization objective regularization. Addressing the instability and vanishing gradient problems common in traditional GAN ​​training, this embodiment employs an improved version of the generative adversarial network (Wasserstein GAN with Gradient Penalty, WGAN-GP), using Earth-Mover distance as the evaluation metric and replacing the original weight pruning mechanism with a gradient penalty term.

[0041] like Figure 1As shown, the DGP framework proposed in this embodiment is a training enhancement structure based on the generative-adversarial approach, consisting of an auxiliary trainer and a master prediction model, where the master prediction model can be arbitrarily replaced with other prediction models. In this framework, the generator, based on GAN adversarial training, further incorporates gradient information from the master prediction model for back-guided fine-tuning, generating pseudo-samples that are more conducive to source-load trend modeling, thereby enhancing the diversity of training data. The discriminator regularizes the master prediction model at the sequence level, providing discriminative feedback at the global distribution level, guiding it to model the global structure of source-load trends. The master prediction model, as the core of the framework, completes future trend modeling through cooperative adversarial interaction with the auxiliary trainer. In this embodiment, to better adapt to the large number of samples generated by the auxiliary trainer, a master prediction model is further proposed, based on a multi-layered encoder stack to form a shared representation backbone, and combined with a multi-output MTL (multi-task learning) architecture to achieve collaborative modeling and accurate prediction of different variable distributions.

[0042] This embodiment provides a deep generative integrated energy system source-load prediction method, which specifically includes the following steps:

[0043] Step 1: Build a generator based on Centcoder to generate virtual training samples that combine real statistical characteristics with unknown change patterns—especially edge scenarios that conform to the system's operating rules, such as sudden changes in wind and solar power output and extreme load fluctuations.

[0044] Step 101: The Transformer excels at capturing long-distance temporal dependencies thanks to its Multi-Head Self-Attention (MHSA) mechanism, possessing powerful global modeling capabilities and parallel computing advantages. To further enhance its feature extraction capabilities, a one-dimensional convolutional neural network (Conv1D) is introduced as a supplementary module, proposing an encoder—Cencoder. In DGP, the generator can synthesize a large amount of data, which makes it possible to build deeper networks to enhance the model's expressive power. However, deep models are prone to gradient vanishing and training degradation problems. To address this, a multi-level residual connection mechanism is introduced in the Cencoder to enhance gradient fluidity and improve training efficiency. The overall structure of the Cencoder is as follows: Figure 2 As shown, it includes a convolution module, a learnable positional encoder, and an encoder module.

[0045] (1) Convolution module.

[0046] First, the embedding layer takes the input sequence Mapping to fixed dimensions of the model dm The embedding representation matrix is ​​obtained as follows:

[0047] (1);

[0048] in, B , L and P These represent the batch size, input time step, and number of features, respectively.

[0049] Subsequently, X embed Processed in parallel via two branches:

[0050] Branch 1 uses Conv1D and MaxPool1D to extract coupling features between different sequences, and stabilizes the distribution by normalizing Norm. The output can be represented as:

[0051] (2);

[0052] It is important to note that X embed Before inputting the convolutional feature extraction module, the dimensions should be converted to B×C×L, where C is the number of channels. The dimensions should be restored after the convolution operation is completed.

[0053] Branch 2 directly applies a one-dimensional convolution to the embedding representation without pooling to preserve local dependencies and fine-grained dynamic changes between adjacent time steps. The output can be represented as:

[0054] (3);

[0055] Finally, to preserve the original features of the data, the original embedding and the outputs of the two branches are added together, and then normalized using LayerNorm to obtain the output of the convolutional module:

[0056] (4);

[0057] (5);

[0058] in, and For learnable parameters, and For the sample mean and standard deviation, It is a local minimum.

[0059] (2) Learnable positional encoding.

[0060] Considering the insensitivity of self-attention mechanisms to sequence order, in X convInjecting positional information is necessary. However, traditional positional encoding is mainly used to express the sequential relationship between positions in a sequence, but... X conv It not only contains sequence data but also other high-dimensional local feature representations, exhibiting more complex spatial-temporal distribution characteristics. Therefore, learnable positional encoding is employed, allowing the model to automatically adjust the representation at each position based on training, thereby improving its ability to model multi-scale features. The output of the learnable positional encoding can be represented as:

[0061] X pos = X conv +P (6);

[0062] in, It is a trainable position encoding matrix, with each time step corresponding to a learnable position vector.

[0063] (3) Encoder module.

[0064] After adding position encoding X pos Long-term dependency feature extraction is performed in the Encoder. Each Encoder layer contains two sub-layers: MHSA (Multi-head Attention Layer) and a feedforward network (Pre-feedback Layer). Residual connections and layer normalization are used after each sub-layer.

[0065] The core of MHSA is the scaled dot product attention formula:

[0066] (7);

[0067] Where softmax is the normalization exponential function. d k Let be the dimension of the key vector. The scaling factor is used to alleviate the gradient vanishing problem caused by excessively large dot product values; Q , K , V These are query, key, and value matrices, respectively, all composed of... X pos The linear transformation is essentially to allow the model to learn the correlation between positions in the sequence from different perspectives.

[0068] Q=X pos W Q ; K=X pos W K ; V=X pos WV (8);

[0069] in, W Q , W K , W K The projection matrix is ​​learnable. MHSA is performed in parallel on multiple independent sets. Q , K , V The projection and attention calculations are performed, and then the outputs of each head are concatenated and linearly transformed to obtain the final output. This allows the model to focus on the feature correlations of different subspaces in the sequence, thereby improving its representational power.

[0070] Definition of the first i The output of the self-attention function is: Then the MSHA output is:

[0071] (9);

[0072] in, h For the number of heads; To output the projection matrix, the multi-head result after the concatenation operation is mapped back to the original dimension; Q i , K i , V i These are the query, key, and value matrices for the i-th self-attention.

[0073] After residual connection and layer normalization, the MSHA sublayer output can be obtained:

[0074] (10);

[0075] Subsequently, The feedforward network typically consists of two fully connected layers, resulting in the feedforward network output:

[0076] (11);

[0077] Where ReLU is a linear rectified function. and This is the weight matrix. and For bias terms, This is the intermediate layer dimension.

[0078] Finally, after residual connection and layer normalization, the Centcoder output is obtained:

[0079] (12);

[0080] Step 102, Generator.

[0081] The generator, built on the Centcoder, effectively models the global temporal dependencies and local coupling features in the source-load sequence, thereby generating virtual training samples that possess both realistic statistical characteristics and unknown change patterns—especially for edge scenarios such as sudden changes in wind and solar power output and extreme load fluctuations, which conform to the system's operating rules. To ensure data controllability, this embodiment employs a conditional generation strategy, using historical data as generation conditions, which, along with a noise vector, are input into the generator network.

[0082] The generation process can be represented as:

[0083] (13);

[0084] in, This is a noise vector sampled from a standard Gaussian distribution; This indicates conditional input, typically real source-load data. x real ; Representing a generator network, such as Figure 1 As shown, it consists of a encoder and a residual structure; This is a generated virtual sample.

[0085] The goal of the generator is to generate samples that can "fool" the discriminator, i.e., minimize the score of the generated samples. The loss function is:

[0086] (14);

[0087] in, P real For the true data distribution in the training set, For generator output, This is the output of the discriminator.

[0088] It is important to note that since the prediction task is a supervised learning activity, the samples generated by the generator consist of the predicted input and output labels.

[0089] Step 2: Construct a discriminator based on a one-dimensional convolutional network. The discriminator acts as an adversarial opponent to the generator, guiding the generator to learn and generate high-quality virtual samples through adversarial training, thereby improving its ability to distinguish between real and generated samples. The generator and discriminator together form an auxiliary trainer.

[0090] The discriminator not only acts as an adversarial opponent to the generator, improving the realism of generated samples through a game-like process, but also serves as an auxiliary structure, providing regularization feedback on the sample distribution to the main prediction model. It's important to note that the discriminator's parameters are updated only during its adversarial training phase with the generator, remaining frozen during the main model training process to avoid introducing additional gradient interference.

[0091] In the discriminator, the input first goes through a combination of one-dimensional convolutional layers, pooling layers, and normalization layers; then, through a flattening layer, the multi-dimensional features are converted into one-dimensional feature vectors; finally, through a Linear layer and a Sigmoid layer, the Sigmoid function maps the output value to between 0 and 1, thereby determining whether the input data is true (real data) or false (fake data generated by the generator).

[0092] The discrimination process can be represented as:

[0093] (15);

[0094] in, x The input samples are used as the generated samples during the adversarial training phase. x gen The prediction result is generated during the training phase of the main prediction model. The core structure is CNN.

[0095] The goal of the discriminator is to distinguish between real data and input data. Its adversarial loss function adopts the WGAN-GP form and is defined as follows:

[0096] (16);

[0097] in, P gen and P interp These are the generation distribution and the interpolation distribution, respectively. The gradient penalty coefficient is... Is x real and x gen Samples obtained by interpolation between them The gradient magnitude of the discriminator with respect to the input is used to constrain the Lipschitz condition.

[0098] Step 3: Introduce the master prediction model. As the core component of the framework, the master prediction model, through cooperation and adversarial interaction with the auxiliary trainer, extracts mapping features from the input data to complete future trend modeling. The master prediction model can be arbitrarily replaced with other existing prediction models.

[0099] Thanks to the large number of pseudo-samples generated by the generator, the main prediction model has stronger data support during training. This allows for the use of deeper network structures to extract more expressive deep features without causing overfitting. Specifically, the input sequence first extracts deep shared representations through a multi-layered stacked encoder, forming a feature representation space with rich semantics. The shared representations at the last time step are used to construct globally shared features, serving as a common basis for the outputs of multiple tasks.

[0100] Let the input be After being processed by multiple layers of encoders, the output is:

[0101] (17);

[0102] Shared features can be represented as:

[0103] (18);

[0104] Considering that different prediction targets have different distribution characteristics and requirements, an independent prediction head is designed for the specific prediction task in the output layer. Shared features are mapped to their respective target output spaces, thereby achieving differentiated feature representation. The loss function needs to simultaneously measure the error of each prediction head and guide the model to maintain the overall prediction accuracy while also considering the balance between different variables.

[0105] The main loss function is defined as a weighted multi-task loss:

[0106] (19);

[0107] in, w i These are the weighting coefficients; It is a loss function used to measure the predicted value in the i-th task. Compared with the true value The error between them; the weights of each prediction task are preset as follows: electrical load 0.4, thermal load 0.3, photovoltaic output 0.15, and wind power output 0.15. The Huber loss combines the advantages of mean absolute error and mean squared error, and is robust to outliers, and can be expressed as:

[0108] (20);

[0109] Among them, threshold Take 1.

[0110] Based on the main regression loss function, this embodiment constructs an adversarial regularization loss based on the discriminator. The aim is to model the global structure and deeper levels of the input data. By using adversarial regularization loss, the model can not only fit the training data more accurately, but also improve its adaptability to unknown data.

[0111] The total loss function is:

[0112] (twenty one);

[0113] in, To combat weight loss and avoid update direction conflicts, this embodiment employs a dynamic weight adjustment strategy based on gradient direction similarity, which can be expressed as:

[0114] (twenty two);

[0115] (twenty three);

[0116] in, , and Do not assign weights to the maximum, minimum, and baseline adversarial losses. This represents gradient direction similarity.

[0117] Step 4: Merge the virtual samples generated by the generator with real historical data to enhance the data diversity of the training set, and input them into the main prediction model for training.

[0118] Step 5: The input first passes through the multi-task learning layer of the main prediction model to learn shared parameters among different sub-tasks, achieving multi-task collaborative optimization. Then, it enters the feature interpretation layer, where multiple prediction heads map the shared features to their respective target output spaces, thereby achieving differentiated feature representation.

[0119] Step 6: The output information of the main prediction model is fed back to the generator, guiding the generator to fine-tune its generation strategy in a secondary manner, so that the output samples are more conducive to the learning of the main task, thereby achieving the collaborative optimization of sample generation and feature modeling.

[0120] The second-order fine-tuning loss function is as follows:

[0121] (twenty four);

[0122] in, and These are the weighting coefficients for the main task and the adversarial losses, respectively.

[0123] Meanwhile, the main prediction model outputs an input discriminator, which is used to determine the authenticity of the prediction results and guide the training of the main prediction model.

[0124] This embodiment selects an IES project in a certain industrial park to validate the proposed DGP, which includes time series data for electrical load, thermal load, photovoltaic power generation, wind power generation, average temperature, wind speed, and irradiance. The dataset contains hourly source-load and meteorological measurement data for 1825 days, ranging from January 1, 2015 to December 31, 2019. To comprehensively evaluate the performance of DGP in the source-load prediction task, four evaluation metrics are used: root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). These metrics cover multiple dimensions such as error magnitude and fitting ability, and can comprehensively judge the model performance from different perspectives.

[0125] To evaluate the effectiveness of the proposed DGP on 24-168h continuous prediction tasks, nine state-of-the-art models were selected as benchmarks for comparison. These benchmark models include not only traditional machine learning methods such as SVR (Support Vector Regression), but also state-of-the-art prediction models based on improved Transformers, such as Informer, Preformer, and Autoformer. Two classic prediction models based on RNNs (Recurrent Neural Networks) – LSTM (Long Short-Term Memory Network) and TCN (Temporal Convolutional Network) – were also considered. Furthermore, commonly used prediction models based on convolutional structures, such as CNN-LSTM (Convolutional Neural Network-Long Short-Term Memory Network) and AGCRN (Graph Attention Recurrent Neural Network) and T-GCN (Temporal Graph Convolutional Network), were also introduced. It should be noted that the above structures are only the backbone; the benchmark models actually include auxiliary structures such as attention mechanisms and multi-task learning. The hyperparameters of the benchmark models were determined using a grid search method.

[0126] Table 1. Results of the comparative experiment

[0127]

[0128] Table 1 shows that the proposed DGP exhibits significant performance advantages in 24-hour and 168-hour continuous forecasting tasks for electricity, heat load, wind, and solar power. Taking electricity load as an example, in the 24-hour continuous forecasting task, compared to the best-performing baseline model, Preformer, DGP reduces MAPE and RMSE by 36.4% and 37.6%, respectively. Its R² is also significantly better than Autoformer's 0.9287. In the 168-hour forecasting task, compared to the best-performing baseline model, Preformer, DGP reduces MAPE and RMSE by 34.55% and 55.86%, respectively, and its R² is also significantly better than Preformer's 0.8423. More importantly, compared to the baseline model where errors rapidly increase in long-term continuous forecasting tasks, DGP demonstrates significant stability and resistance to error propagation, showing the smallest accuracy difference in both 24-hour and 168-hour continuous forecasting tasks. Compared to traditional LSTM, CNN-LSTM achieves higher accuracy, indicating that adding convolutional modules to the recurrent structure effectively enhances the model's ability to extract local features. This is one of the inspirations for designing the Centcoder in this embodiment. Among all baseline models, the Transformer series models generally outperform traditional prediction models, demonstrating significant advantages in complex source-load time-series tasks. To more intuitively illustrate the prediction results, some representative models are selected for visualization, and actual data is provided, such as... Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown.

[0129] Example 2

[0130] This embodiment provides a deep generative integrated energy system source-load prediction system, which specifically includes:

[0131] The training module is configured as follows: based on real samples in the training set, the discriminator acts as the adversarial object of the generator, and the generator learns to generate virtual samples through adversarial training; after merging real samples and virtual samples, the results are input into the main prediction model, and the output of the main prediction model is input into the discriminator. The discriminator is used to judge the authenticity of the prediction results and guide the training of the main prediction model.

[0132] The data acquisition module is configured to acquire source-load data and meteorological data of the integrated energy system at several consecutive historical moments.

[0133] The prediction module is configured to obtain future source-load data based on source-load data and meteorological data from several consecutive historical moments, using the master prediction model.

[0134] Furthermore, both the generator and the main prediction model employ an encoder. The encoder uses an embedding layer to map the input to a fixed dimension, and then processes it in parallel through two branches. The output of the embedding layer and the output of the two branches are added together and then passed through layer normalization, learnable position encoding, multi-head attention layer and pre-feedback layer in sequence to obtain the encoder output. Residual connections and layer normalization are used after the multi-head attention layer and the pre-feedback layer.

[0135] Furthermore, the main prediction model is trained using a weighted sum of weighted multi-task loss and adversarial regularization loss, with the weights of the adversarial regularization loss being: ;in, , and These are the maximum adversarial loss weight, the minimum adversarial loss weight, and the baseline adversarial loss weight, respectively. For gradient direction similarity, , To weighted multi-task loss, To counteract regularization loss.

[0136] Furthermore, the source-load data includes electrical load, thermal load, photovoltaic power generation, and wind power generation.

[0137] Furthermore, the meteorological data includes average temperature, wind speed, and irradiance.

[0138] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0140] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep generative integrated energy system source-load prediction method, characterized in that, The method comprises: obtaining source-load data and meteorological data of a comprehensive energy system at a plurality of continuous historical time points; obtaining source-load data at a future time point based on the source-load data and meteorological data at the plurality of continuous historical time points through a main prediction model; wherein the training step of the main prediction model comprises: taking real samples in a training set as the adversarial object of a generator, guiding the generator to learn to generate virtual samples through adversarial training; and inputting the real samples and the virtual samples into the main prediction model, outputting the main prediction model into a discriminator, judging the authenticity of the prediction result by the discriminator, and guiding the training of the main prediction model. The main prediction model training adopts a weighted sum of a weighted multi-task loss and an adversarial regularization loss, and a weight of the adversarial regularization loss is: ; wherein, , and are a maximum adversarial loss weight, a minimum adversarial loss weight, and a benchmark adversarial loss weight, respectively, is a gradient direction similarity, , is a weighted multi-task loss, is an adversarial regularization loss.

2. The deep generative integrated energy system source-load forecasting method of claim 1, wherein, The generator and the main prediction model both adopt an encoder, and the encoder adopts an embedding layer to map the input to a fixed dimension, then processes the embedding layer output and the two branch outputs in parallel, adds the outputs, and then sequentially passes through layer normalization, learnable position encoding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and residual connection and layer normalization are used after the multi-head attention layer and the pre-feedback layer.

3. The deep generative integrated energy system source-load forecasting method of claim 1, wherein, The source-load data includes electrical load, thermal load, photovoltaic power generation and wind power generation.

4. The deep generative integrated energy system source-load forecasting method of claim 1, wherein, The meteorological data includes average temperature, wind speed and irradiance.

5. A deep generative integrated energy system source-load forecasting system, characterized in that, The method comprises: a training module configured to take real samples in a training set as the adversarial object of a generator, guide the generator to learn to generate virtual samples through adversarial training, input the real samples and the virtual samples into a main prediction model, output the main prediction model into a discriminator, judge the authenticity of the prediction result by the discriminator, and guide the training of the main prediction model; The main prediction model training adopts a weighted sum of a weighted multi-task loss and an adversarial regularization loss, and a weight of the adversarial regularization loss is: ; wherein, , and are a maximum adversarial loss weight, a minimum adversarial loss weight, and a benchmark adversarial loss weight, respectively, is a gradient direction similarity, , is a weighted multi-task loss, is an adversarial regularization loss. a data acquisition module configured to obtain source-load data and meteorological data of a comprehensive energy system at a plurality of continuous historical time points; a prediction module configured to obtain source-load data at a future time point based on the source-load data and meteorological data at the plurality of continuous historical time points through a main prediction model.

6. A deep generative integrated energy system source- demand forecasting system as claimed in claim 5, wherein, The generator and the main prediction model both adopt an encoder, and the encoder adopts an embedding layer to map the input to a fixed dimension, then processes the embedding layer output and the two branch outputs in parallel, adds the outputs, and then sequentially passes through layer normalization, learnable position encoding, a multi-head attention layer and a pre-feedback layer to obtain the encoder output, and residual connection and layer normalization are used after the multi-head attention layer and the pre-feedback layer.

7. A deep generative integrated energy system source- demand forecasting system as claimed in claim 5, wherein, The source-load data includes electrical load, thermal load, photovoltaic power generation and wind power generation.

8. The deep generative integrated energy system source- demand forecasting system of claim 5, wherein, The meteorological data includes average temperature, wind speed and irradiance.

Images