A method and system for dynamic balance management of enterprise energy supply and demand

Abstract

This invention relates to the field of energy management technology, and more particularly to a method and system for dynamic balance management of enterprise energy supply and demand. The method collects multi-source heterogeneous data from enterprises; extracts key features through spatiotemporal alignment and data fusion, and utilizes a multimodal variational autoencoder; inputs these features and enterprise physical topology parameters into an energy supply and demand balance prediction model constructed by fusing a deep kernel learning neural network based on kernel function improvement and a deep Gaussian process, predicting net energy demand for future periods; and then inputs the prediction results into an energy dispatch optimization decision model constructed based on a dense connection network and attention mechanism to generate the optimal energy allocation strategy. This method achieves closed-loop management of deep multi-source data fusion, accurate demand prediction, and intelligent optimization dispatch, improving the operational efficiency, economy, and dynamic response capability of enterprise energy systems, and is suitable for multi-energy collaborative management in complex industrial scenarios.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a method and system for dynamic balance management of enterprise energy supply and demand. Background Technology

[0002] As industrial energy systems rapidly evolve towards intelligence and low carbon emissions, enterprises face increasingly complex challenges in coordinating energy supply and demand during production and operation. On one hand, internal energy consumption within enterprises is characterized by multi-energy complementarity and dynamic changes, involving various energy forms such as electricity, natural gas, and heat, and is significantly affected by the operating status of production equipment, the charging and discharging strategies of energy storage systems, and fluctuations in the output of distributed renewable energy sources (such as photovoltaics and wind power). On the other hand, external environmental factors such as grid electricity price signals, weather conditions, and adjustments to production plans also significantly disrupt energy demand. Traditional energy management methods often rely on manual experience or scheduling strategies based on static models, making it difficult to effectively integrate and dynamically respond to multi-source heterogeneous data. This results in low energy utilization efficiency, high operating costs, and an inability to adapt to the flexible and interactive needs of new power systems.

[0003] Existing energy forecasting and scheduling models have limitations when dealing with high-dimensional, nonlinear, and highly uncertain industrial energy data, particularly in terms of data fusion capabilities, accuracy of time-series feature extraction, and long-term forecast stability. Furthermore, most scheduling optimization methods lack the ability to jointly model the system's physical topology and real-time operating status, making it difficult to generate optimal energy allocation schemes that balance economy, reliability, and sustainability.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for dynamic balance management of enterprise energy supply and demand, which aims to solve the technical problem that existing enterprise energy management methods are unable to effectively integrate multi-source heterogeneous data and accurately respond to dynamic changes in energy supply and demand, resulting in low energy utilization efficiency, poor prediction accuracy, weak scheduling and optimization capabilities, and inability to meet the dynamic balance requirements under complex operating conditions.

[0006] To achieve the above objectives, the present invention provides a method for dynamic balance management and control of enterprise energy supply and demand, the method comprising:

[0007] Collect multi-source heterogeneous data from enterprises, including one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data.

[0008] The multi-source heterogeneous data is spatiotemporally aligned and fused, and key features are extracted from the fused data based on a multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data.

[0009] The key features and the enterprise's physical topology parameters are input into a pre-built energy supply and demand balance prediction model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance prediction model. The energy supply and demand balance prediction model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process.

[0010] The net energy demand is input into a pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model. The energy dispatch optimization decision model is built based on a dense connection network combined with an attention mechanism network.

[0011] Optionally, the step of inputting the key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance prediction model to obtain the enterprise's net energy demand for future periods output by the energy supply and demand balance prediction model includes:

[0012] The key features and the enterprise's physical topology parameters are input into a deep kernel learning neural network for deep feature extraction to obtain the deep features of the key features.

[0013] Based on the kernel function, the deep features are mapped to a high-dimensional feature space to obtain high-dimensional deep features;

[0014] Net energy demand is predicted based on the high-dimensional deep features using a deep Gaussian process, yielding the enterprise's net energy demand over a predetermined future time period.

[0015] Optionally, the step of inputting the net energy demand into a pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model includes:

[0016] The net energy demand is input into a dense network for nonlinear feature extraction to obtain the first scheduling feature.

[0017] The net energy demand is input into the attention mechanism network, and the second scheduling feature is extracted by weighting the electricity price fluctuation and the temporal importance.

[0018] The first and second scheduling features are merged to generate a comprehensive scheduling feature. Based on this comprehensive scheduling feature, the optimal energy allocation strategy that satisfies economic, stability and carbon emission constraints is solved, and control commands for each energy unit are output.

[0019] Optionally, the key feature extraction of the fused data based on the multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data includes:

[0020] The encoder maps the fused multi-source heterogeneous data into a probability distribution of latent variables, which represent the implicit operating state of the energy system.

[0021] Based on the decoder, the original multimodal data is reconstructed from the distribution of the latent variables to form a reconstructed data matrix;

[0022] By minimizing the reconstruction error and KL divergence loss, a multimodal variational autoencoder is trained, and key features are extracted from the latent representation of the encoder output.

[0023] Optionally, the training process of the energy supply and demand balance prediction model includes:

[0024] Based on the stochastic gradient Hamiltonian Monte Carlo sampling method, the weight parameters of the initial energy supply and demand balance prediction model are sampled under the Bayesian framework to achieve posterior distribution approximation, complete the preliminary parameter adjustment, and obtain the energy supply and demand balance prediction model after preliminary parameter adjustment.

[0025] Furthermore, stochastic gradient Langevin dynamics is introduced to inject controllable noise during parameter updates, enhancing the model's generalization ability and avoiding getting trapped in local optima. The model parameters are adaptively optimized, ultimately resulting in a robust energy supply and demand balance prediction model.

[0026] Optionally, the training process of the energy dispatch optimization decision model includes:

[0027] During backpropagation, an adaptive gradient pruning mechanism is used to dynamically adjust the gradient pruning threshold by using the second moment estimation of the gradient to prevent gradient explosion, thus obtaining an energy scheduling optimization decision model with preliminary parameter adjustments.

[0028] Furthermore, an adaptive momentum estimation algorithm is adopted to dynamically adjust the learning rate and momentum factor based on the first and second moments of the gradient, thereby improving the convergence speed and stability of the model and finally training a high-performance energy dispatch optimization decision model.

[0029] Furthermore, to achieve the above objectives, the present invention also provides an enterprise energy supply and demand dynamic balance management and control system, the system comprising:

[0030] The data acquisition module is used to collect multi-source heterogeneous data from the enterprise. The multi-source heterogeneous data includes one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data.

[0031] The feature fusion module is used to perform spatiotemporal alignment and data fusion on the multi-source heterogeneous data, and to extract key features from the fused data based on a multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data.

[0032] The demand forecasting module is used to input the key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance forecasting model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance forecasting model. The energy supply and demand balance forecasting model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process.

[0033] The scheduling decision module is used to input the net energy demand into a pre-built energy scheduling optimization decision model to obtain the optimal energy allocation strategy output by the energy scheduling optimization decision model. The energy scheduling optimization decision model is constructed based on a dense connection network combined with an attention mechanism network.

[0034] In addition, to achieve the above objectives, the present invention also provides an enterprise energy supply and demand dynamic balance management and control device, the device comprising: a memory, a processor, and an enterprise energy supply and demand dynamic balance management and control program stored in the memory and executable on the processor, the enterprise energy supply and demand dynamic balance management and control program being configured to implement the steps of the enterprise energy supply and demand dynamic balance management and control method as described above.

[0035] In addition, to achieve the above objectives, the present invention also provides a medium storing an enterprise energy supply and demand dynamic balance management program, wherein when the enterprise energy supply and demand dynamic balance management program is executed by a processor, it implements the steps of the enterprise energy supply and demand dynamic balance management method as described above.

[0036] In addition, to achieve the above objectives, the present invention also provides a computer program product, including an enterprise energy supply and demand dynamic balance management and control program, wherein when the enterprise energy supply and demand dynamic balance management and control program is executed by a processor, it implements the steps of the enterprise energy supply and demand dynamic balance management and control method as described above.

[0037] This invention provides a method for dynamic energy supply and demand balance management in enterprises. By introducing a spatiotemporal alignment mechanism and a multimodal variational autoencoder, this method effectively integrates multi-dimensional and heterogeneous data from within the enterprise, including electricity, gas, heat, equipment status, energy storage, renewable energy, and external data such as electricity prices, weather, and production plans. This overcomes the weakness of traditional methods in processing heterogeneous data, significantly improving the completeness of data representation and the accuracy of key feature extraction. The constructed energy supply and demand balance prediction model integrates a deep kernel learning neural network with improved kernel functions and a deep Gaussian process, possessing both strong nonlinear modeling capabilities and uncertainty quantification capabilities. It can accurately capture the dynamic changing trends of energy demand in complex industrial scenarios, especially in the face of renewable energy fluctuations. The method exhibits higher predictive stability and reliability under dynamic and production uncertainties. The energy dispatch optimization decision model adopts a dense connection network combined with an attention mechanism, which can fully explore the key dependencies in historical and real-time information, adaptively focus on the core factors affecting dispatch effectiveness, and thus generate the optimal energy allocation strategy that takes into account economy, energy efficiency and system security. This method constructs a complete intelligent management and control chain from data collection, feature fusion, demand forecasting to dispatch decision-making. Combined with enterprise physical topology parameter modeling, it realizes global perception and dynamic optimization of the energy system, significantly improves enterprise energy utilization efficiency, reduces energy costs, enhances the ability to respond to internal and external disturbances, and promotes the transformation of enterprises towards green, intelligent and resilient energy management. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating an embodiment of the enterprise energy supply and demand dynamic balance management method of the present invention.

[0039] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0040] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0041] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the enterprise energy supply and demand dynamic balance management method of the present invention, which presents an embodiment of the enterprise energy supply and demand dynamic balance management method of the present invention.

[0042] In one embodiment, the enterprise energy supply and demand dynamic balance management method includes:

[0043] Step S100: Collect multi-source heterogeneous data from the enterprise. The multi-source heterogeneous data includes one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data.

[0044] The multi-source heterogeneous data can be a collection of energy-related data from different systems, with different structures, sampling frequencies, time granularities, and semantic meanings. It can be used as the raw input for the perception layer of an enterprise's energy system, supporting subsequent fusion, prediction, and scheduling decisions. In this embodiment, the multi-source heterogeneous data can include, but is not limited to, one or more of structured time-series data, unstructured text plans, and semi-structured equipment logs.

[0045] Electricity load data can be a record of the electricity consumption of various electrical equipment or areas within an enterprise per unit time. It can be used to reflect the dynamics of electricity demand and is one of the basic inputs for constructing net energy demand. Gas consumption data can be a record of the consumption of natural gas or other gaseous fuels by an enterprise per unit time. It can be used to characterize the gas energy demand in thermal or process operations. Heat consumption data can be a record of the usage of heat energy carriers such as steam and hot water by an enterprise per unit time. It can be used to reflect the heat energy supply and demand status and participate in multi-energy complementarity modeling. Production equipment operation status data can be real-time information reflecting the operating parameters of production equipment such as start-up and shutdown, load rate, and fault status. It can be used to correlate production activities with energy consumption and improve the causal explanatory power of forecasting and scheduling. Energy storage system status data can be real-time operating indicators of energy storage units, including battery state of charge (SOC), charging and discharging power, and available capacity. It can be used to provide information on system flexibility resources and influence the calculation of net energy demand and the generation of scheduling strategies.

[0046] Renewable energy generation data can be actual or projected output data from distributed power sources such as photovoltaics and wind power within an enterprise. This data can be used as a negative load in net energy demand calculations, and its volatility needs to be covered by uncertainty modeling. External grid price signals can be time-of-use prices, real-time prices, or demand response price signals from the electricity market. These can be used to constitute external cost constraints for economic dispatch and drive energy consumption optimization. Meteorological and environmental data can include environmental parameters affecting energy production and consumption, such as temperature, irradiance, wind speed, and humidity. This data can be used to correct renewable energy output forecasts and heat load models. Enterprise production plan data can include production scheduling information such as production arrangements, order tasks, and shift adjustments for future periods. This data can be used as feedforward signals to predict energy demand trends in advance.

[0047] Collecting multi-source heterogeneous data from enterprises can be achieved through various energy-related data sources such as sensors, SCADA systems, ERP interfaces, and meteorological APIs. Furthermore, the collection of multi-source heterogeneous data from enterprises can be realized through multi-channel parallel access, protocol adaptation, and format standardization, thereby building a sensing foundation covering all energy forms and internal and external disturbance factors.

[0048] Step S200: Spatiotemporal alignment and data fusion are performed on the multi-source heterogeneous data, and key features are extracted from the fused data based on the multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data.

[0049] The spatiotemporal alignment mechanism can be a processing flow that maps multi-source data with different timestamps, sampling frequencies, and spatial granularities to a unified spatiotemporal coordinate system. This can be used to eliminate data asynchronicity and scale differences, providing a consistent foundation for subsequent fusion. In an exemplary embodiment, the spatiotemporal alignment mechanism can achieve alignment through interpolation, resampling, time window aggregation, and spatial topology mapping. Furthermore, the spatiotemporal alignment mechanism can collaborate with multi-source heterogeneous data to provide a standardized input format; and collaborate with the data fusion module to ensure that the data is consistent in the spatiotemporal dimension before fusion.

[0050] The data fusion module can be a processing unit that semantically integrates and structurally unifies multi-source data after spatiotemporal alignment, and can be used to generate joint representation tensors that can be processed by deep models. Key features can be low-dimensional, high-information-density, and physically interpretable latent variable representations extracted from the fused data, which can be used as inputs to predictive models to reduce redundancy and enhance generalization ability.

[0051] A multimodal variational autoencoder (VAE) can be a generative model based on variational inference, capable of jointly modeling multimodal input data and learning a shared latent space. It can be used to decouple robust, compact key features from high-dimensional heterogeneous data, suppress noise, and preserve cross-modal associations. For example, a VAE can optimize the encoder-decoder structure by maximizing the lower bound of evidence (ELBO), forcing different modalities to align in the latent space. In one specific embodiment, the VAE can receive the output of a data fusion module, outputting key features for use by a prediction model. Furthermore, a VAE can include, but is not limited to, one or more of conditional multimodal VAEs, cross-modal contrastive VAEs, and hierarchical multimodal VAEs.

[0052] Spatiotemporal alignment and data fusion of multi-source heterogeneous data can be achieved by unifying data with different timestamps and spatial granularities into a standard spatiotemporal grid through interpolation, aggregation, mapping, etc., and then performing semantic integration. Furthermore, spatiotemporal alignment and data fusion of multi-source heterogeneous data can be achieved by using linear / spline interpolation to upsample low-frequency data and moving average downsampling to high-frequency data, as well as spatial alignment based on device ID or region coding; or by constructing a unified time index table, linking multi-source data tables through foreign keys, and then performing tensor quantization and concatenation, thereby eliminating data asynchronicity and structural fragmentation, and forming a structured joint input.

[0053] Extracting key features from fused data using a multimodal variational autoencoder (VAE) can be achieved by inputting the fused data into a multimodal VAE, mapping it to a shared latent space through the encoder, and outputting low-dimensional latent variables as key features. Furthermore, key feature extraction from fused data based on a VAE can be achieved by employing a modality-specific encoder with a shared latent layer structure, aligning the distribution through KL divergence regularization; or by introducing a cross-modal reconstruction loss to force different modalities to maintain semantic consistency in the latent space. This allows for the extraction of robust, decoupled, and physically meaningful feature representations from high-dimensional noisy data.

[0054] Step S300: Input the key features and the enterprise's physical topology parameters into the pre-built energy supply and demand balance prediction model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance prediction model. The energy supply and demand balance prediction model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process.

[0055] The enterprise physical topology parameters can be a set of parameters describing the physical structure of the enterprise's energy system equipment, energy flow paths, capacity limitations, etc., and can be used to provide physical constraint priors for prediction and scheduling models to ensure the feasibility of the solution. In a specific embodiment, the enterprise physical topology parameters may include, but are not limited to, one or more of the following: pipeline topology matrix, electrical connection diagram, heat transfer impedance parameters, etc.

[0056] Net energy demand can be defined as the net load of a company's total energy demand minus the output of local renewable energy sources over a future period. Decomposed by energy form, it can be used as input to a scheduling optimization model to drive the generation of the optimal allocation strategy. The energy supply and demand balance forecasting model can be a composite machine learning model used to predict a company's future net energy demand, providing high-precision energy demand forecasts with uncertainty quantification. Furthermore, the energy supply and demand balance forecasting model can receive key features and the company's physical topology parameters to output net energy demand.

[0057] Deep kernel learning neural networks with improved kernel functions can be model structures that introduce learnable kernel functions into the output layer of a deep neural network to enhance its nonlinear mapping capabilities. This can be used to improve the fitting ability to strongly nonlinear dynamics of energy systems. For example, a deep kernel learning neural network with improved kernel functions can embed the neural network into the covariance function of a Gaussian process, learning complex dependencies through the kernel function parameters. In an exemplary embodiment, the deep kernel learning neural network with improved kernel functions can include, but is not limited to, one or more of spectral kernel deep learning, multi-scale kernel networks, and adaptive radial basis function kernel networks.

[0058] A deep Gaussian process can be a probabilistic generative model composed of stacked layers of Gaussian processes. It possesses hierarchical uncertainty propagation capabilities and can be used to quantify the confidence interval of prediction results, supporting risk-aware decision-making. Furthermore, the deep Gaussian process can approximate the posterior distribution through stochastic variational inference or Monte Carlo sampling. In a specific embodiment, the deep Gaussian process can include, but is not limited to, one or more of sparse deep Gaussian processes, convolutional deep Gaussian processes, and graph-structured deep Gaussian processes.

[0059] By inputting key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance forecasting model, or by concatenating or embedding key feature vectors with physical topology parameters (as shown in the diagram or constraint matrix) before inputting them into the forecasting model, data-driven approaches and physical priors can be integrated to improve the accuracy and feasibility of forecasts. The net energy demand of the enterprise in future time periods is obtained from the energy supply and demand balance forecasting model. This can be achieved by forward propagation of the model to generate net energy demand forecasts and their uncertainty ranges for multiple future time steps, thus providing confident multi-energy net load forecasts to support risk-controlled scheduling decisions.

[0060] Step S400: Input the net energy demand into the pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model. The energy dispatch optimization decision model is built based on a dense connection network combined with an attention mechanism network.

[0061] The optimal energy allocation strategy can be a comprehensive energy procurement, conversion, storage, and usage scheme that achieves the best overall performance in terms of economy and energy efficiency while meeting physical constraints and operational safety. It can be used to directly guide the real-time operation and control of an enterprise's energy system. The energy dispatch optimization decision model can be a deep optimization network that generates the optimal energy allocation strategy based on prediction results. It can be used to achieve multi-objective collaborative optimization of economy, energy efficiency, and safety. Furthermore, the energy dispatch optimization decision model can receive net energy demand and output the optimal energy allocation strategy. The dense connection network can be a deep neural network architecture where each layer is directly connected to all subsequent layers. It can be used to enhance the model's efficiency in utilizing multi-scale historical information. For example, the dense connection network can concatenate shallow and deep features through skip connections, mitigating gradient vanishing and promoting feature reuse. In a specific embodiment, the dense connection network can include, but is not limited to, one or more of the following: DenseNet variants, cross-stage dense blocks, and channel-level dense connections.

[0062] Attention mechanism networks can be neural network components that dynamically assign importance to elements in an input sequence using learnable weights. They can adaptively focus on the state variables or time steps that have the greatest impact on the scheduling objective. Furthermore, attention mechanism networks can generate weighted context vectors by calculating the similarity scores of query-key-value triples. In an exemplary embodiment, the attention mechanism network can include, but is not limited to, one or more of multi-head attention, temporal attention, and channel attention. Inputting the net energy demand into a pre-built energy scheduling optimization decision model can involve feeding the predicted net demand sequence as the input sequence to the scheduling model, triggering a strategy generation process, thereby initiating the optimization solution process and generating feasible scheduling schemes. Obtaining the optimal energy allocation strategy output by the energy scheduling optimization decision model can involve the model outputting control commands for the start-up and shutdown, power setting, and energy storage charging and discharging of various energy devices in future time periods, thereby generating an executable energy scheduling scheme that balances economy, energy efficiency, and system safety.

[0063] Taking multi-energy collaborative scheduling in an industrial park as an example, the enterprise energy supply and demand dynamic balance management method in this embodiment can be as follows: A chemical industrial park deploys this method: In stage S100, it simultaneously collects the electrical load of 200 motors, the heat consumption of 3 steam pipelines, the output of 5MW rooftop photovoltaic power, the daily time-of-use electricity price, the weather forecast for the next 72 hours, and the production schedule for the following week; In stage S200, it unifies the second-level equipment data and hourly-level meteorological data to a 15-minute granularity through a spatiotemporal alignment mechanism and merges them into a multimodal tensor; In stage S300, a multimodal VAE... 128 key features were extracted and combined with the physical topology parameters of the cogeneration units and energy storage stations in the park. The data were then input into the energy supply and demand balance prediction model, which output the net demand for electricity and steam for the next 24 hours and the 95% confidence interval. Based on this prediction, the S400 stage scheduling model reused historical scheduling experience through a dense network and identified the current high electricity price period and the low photovoltaic output period as key risk points through an attention mechanism. Finally, the optimal strategy of prioritizing the use of energy storage and delaying the operation of non-critical production lines was generated, thereby reducing energy costs and carbon emissions.

[0064] This embodiment provides a method for dynamic balance management of enterprise energy supply and demand. It collects multi-source heterogeneous data from enterprises, performs spatiotemporal alignment and data fusion on this data, and extracts key features from the fused data based on a multimodal variational autoencoder. These key features, along with the enterprise's physical topology parameters, are input into a pre-built energy supply and demand balance prediction model, while net energy demand is input into a pre-built energy dispatch optimization decision model. By integrating multi-dimensional information such as electricity, natural gas, heat, equipment status, energy storage, renewable energy, and external electricity prices, meteorology, and production plans, it addresses the perception blind spots caused by data silos and heterogeneity. Spatiotemporal alignment and data fusion ensure that data with different sampling frequencies, timestamps, and spatial granularities are collaboratively modeled under a unified spatiotemporal benchmark, thereby improving data efficiency. Consistency is achieved by decoupling physically meaningful low-dimensional representations from high-dimensional noisy data using a multimodal variational autoencoder to enhance feature robustness; by fusing key features and physical topology parameters and utilizing a deep kernel learning neural network improved by kernel functions to enhance nonlinear fitting capabilities; by quantifying prediction uncertainty through a deep Gaussian process to improve prediction accuracy and reliability; by promoting multi-scale information reuse through dense connection networks and dynamically focusing key state variables through attention mechanisms to achieve adaptive optimization scheduling. This can form a closed-loop intelligent control chain from perception and prediction to decision-making. Combined with physical topology constraints, the system possesses global coordination, dynamic response, and anti-disturbance capabilities, significantly improving energy utilization efficiency, reducing energy costs, and supporting enterprises' transformation towards green, intelligent, and resilient energy management.

[0065] In one embodiment, key features and the enterprise's physical topology parameters are input into a pre-built energy supply and demand balance forecasting model to obtain the enterprise's net energy demand for future periods, as output by the energy supply and demand balance forecasting model, including:

[0066] The key features and the enterprise's physical topology parameters are input into a deep kernel learning neural network to extract deep features, thus obtaining the deep features of the key features.

[0067] The deep kernel learning neural network can be a composite model structure that embeds a deep neural network into the covariance function of a Gaussian process and extracts nonlinear features through a learnable kernel function. It can be used to automatically mine higher-order nonlinearities from key features and physical topological parameters, generating more discriminative deep features. In this embodiment, the deep kernel learning neural network uses a neural network as the feature mapping function, and its output is used to construct the kernel matrix of the Gaussian process, and is jointly optimized through end-to-end training. Furthermore, the deep kernel learning neural network can include, but is not limited to, one or more of spectral normalization deep kernel networks, multi-task deep kernel networks, and graph-aware deep kernel networks. Deep features can be abstract representations output from the intermediate layers of the deep kernel learning neural network, integrating the original key features and physical topological information. They can be used as input to the kernel mapping, preserving the deep nonlinear structure and providing a foundation for subsequent high-dimensional mapping.

[0068] Key features and the enterprise's physical topology parameters are input into a deep kernel learning neural network for deep feature extraction. This can be achieved by concatenating or embedding key feature vectors with physical topology parameters (such as adjacency matrices or constraint vectors) before inputting them into the deep kernel learning neural network, generating deep features through forward propagation. Furthermore, this operation can be achieved by fusing data-driven features with prior knowledge of physical structures, thereby enhancing the model's interpretability and generalization ability regarding energy system dynamics. The resulting deep features can be extracted from the hidden or output layers of the deep kernel learning neural network. Furthermore, this operation allows for the acquisition of compact representations containing high-order nonlinear interaction information, which is superior to a linear combination of the original key features.

[0069] Based on kernel functions, deep features are mapped to a high-dimensional feature space to obtain high-dimensional deep features.

[0070] The kernel function can be a positive definite function used to measure the similarity of two inputs in a high-dimensional (potentially infinite-dimensional) regenerative kernel Hilbert space. It can be used to implicitly implement nonlinear mappings, avoid explicit computation of high-dimensional coordinates, and improve the ability to model complex energy coupling relationships. In an exemplary embodiment, the kernel function can employ an Automatic Relevance Determination (ARD) kernel to assign learnable weights to different feature dimensions; or it can use a composite kernel (such as RBF + periodic kernel) from a deep kernel learning framework to capture multi-scale energy dynamics. Exemplarily, the kernel function can include, but is not limited to, one or more of radial basis function kernels, Marton kernels, and periodic kernels.

[0071] A high-dimensional feature space can be a regenerating kernel Hilbert space induced by a kernel function, where the original features are nonlinearly mapped to a higher-dimensional representation. This can be used to make energy patterns that are originally inseparable in low-dimensional space linearly separable or more strongly correlated in high-dimensional space. In a specific embodiment, the high-dimensional feature space may include, but is not limited to, the Gaussian process feature space, the Mercer kernel space, the implicit tensor product space, etc.

[0072] High-dimensional deep features can be the equivalent representation of deep features in a high-dimensional feature space after implicit mapping by a kernel function. They can be used as input to deep Gaussian processes to support high-precision, uncertainty-aware net energy demand forecasting.

[0073] Based on kernel functions, mapping deep features to a high-dimensional feature space can be achieved without explicitly calculating the mapping result, but rather by calculating the inner product of any two deep features in the high-dimensional space using a kernel function. Furthermore, this operation can be implemented using an Automatic Relevance Determination (ARD) kernel or a composite kernel within a deep kernel learning framework, thereby effectively modeling complex nonlinear coupling relationships such as wind and solar power output and load response. Obtaining high-dimensional deep features can be achieved by implicitly holding the geometric relationships of high-dimensional deep features in the form of a kernel matrix within the kernel method framework. Furthermore, this operation can provide a highly expressive input representation for deep Gaussian processes.

[0074] Net energy demand is predicted based on high-dimensional deep features using deep Gaussian processes, yielding the net energy demand of enterprises within a preset future time period.

[0075] The enterprise's future preset time period can be a time window for the scheduling system to make predictions and decisions, typically a discrete time series of several hours to several days. This can be used to define the prediction range of net energy demand and match it with the scheduling cycle. Net energy demand prediction based on high-dimensional deep features using a deep Gaussian process can be achieved by taking high-dimensional deep features as input, modeling the output distribution through the hierarchical covariance structure of the deep Gaussian process, and outputting the predicted mean and variance. Furthermore, this operation can be achieved by using stochastic variational inference to process large-scale time series data, introducing induced point approximate posteriors, or constructing a multi-output deep Gaussian process to simultaneously predict the net demand of multiple energy sources such as electricity, heat, and gas. This allows for the provision of probabilistic prediction results while maintaining high fitting accuracy, supporting risk-aware scheduling decisions. The net energy demand of the enterprise within the future preset time period can be obtained by forward propagation of the deep Gaussian process to generate predicted net energy demand values ​​and their uncertainty intervals for multiple future time steps. Furthermore, this operation can output a multi-energy net load sequence with confidence, serving as a reliable input to the scheduling optimization model.

[0076] Taking the forecasting of high-proportion renewable energy parks as an example, the enterprise energy supply and demand dynamic balance management method in this embodiment can be as follows: A manufacturing park deploys this forecasting process, the key features of which include photovoltaic output fluctuations, production line start-up and shutdown signals and electricity price signals, and physical topology parameters describe energy storage and cogeneration units; a deep kernel learning neural network extracts the deep features of the coupling constraints of the integrated equipment; a kernel function (such as RBF + periodic composite kernel) maps it to a high-dimensional space to capture the nonlinear response under daily cycles and weather changes; a deep Gaussian process predicts the net electricity demand for the next 24 hours based on this representation, and outputs the mean curve and 95% confidence band; when there is a sudden change in cloud cover in the forecast for the next day, the model automatically expands the uncertainty interval, and the dispatching system reserves more reserve capacity accordingly to avoid the risk of load shedding.

[0077] This embodiment provides a method for dynamic balance management of enterprise energy supply and demand. It extracts deep features by inputting key characteristics and the enterprise's physical topology parameters into a deep kernel learning neural network. The deep structure automatically mines high-order nonlinear interaction patterns in the original features and integrates system physical constraints, enhancing the model's physical interpretability and generalization ability. By mapping deep features to a high-dimensional feature space based on kernel functions, and implicitly constructing high-dimensional or even infinite-dimensional representations using kernel tricks, it significantly improves the ability to model complex energy coupling relationships. Furthermore, by using a deep Gaussian process to predict net energy demand from high-dimensional deep features, it inherits the probabilistic output capability of Gaussian processes for predictive uncertainty and handles large-scale high-dimensional inputs. This allows for high-precision net energy demand prediction with confidence intervals even under disturbances such as drastic fluctuations in renewable energy or sudden changes in production plans. It achieves integrated modeling from physical perception and nonlinear representation to uncertainty quantification, fundamentally addressing the shortcomings of traditional methods in long-term prediction stability, nonlinear fitting ability, and risk perception. This provides reliable and robust input for scheduling decisions, ultimately supporting the generation of energy allocation strategies that balance economy, reliability, and sustainability.

[0078] In one embodiment, the net energy demand is input into a pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model, including:

[0079] The net energy demand is input into a densely connected network for nonlinear feature extraction to obtain the first scheduling feature.

[0080] The nonlinear features can be high-order interaction patterns or implicit structural information extracted from the original input through nonlinear transformation, and can be used to characterize the complex nonlinear dependency between energy demand and system response. In this embodiment, the nonlinear features are obtained by performing layer-by-layer nonlinear mapping on the net energy demand sequence through a densely connected network. The first scheduling feature can be an intermediate representation extracted from the net energy demand by the densely connected network, retaining the multi-scale nonlinear structure, and can be used as a structural feature to support subsequent fusion, enhancing the scheduling model's ability to model system dynamics. Further, inputting the net energy demand into the densely connected network for nonlinear feature extraction to obtain the first scheduling feature can be achieved by feeding the net energy demand sequence into the densely connected network and using its fully connected inter-layer structure to extract multi-scale nonlinear patterns. For example, this operation can be achieved by using a DenseBlock structure, concatenating the output of each layer to the input of all subsequent layers, and forming deep features after nonlinear activation; or by introducing global average pooling at the end of the network to compress the temporal dimension and retain channel-level nonlinear interaction information, thereby preserving the complex dynamic structure in energy demand and generating robust features sensitive to system response.

[0081] The net energy demand is input into the attention mechanism network, and the second scheduling feature is extracted by weighting the electricity price fluctuations and the temporal importance.

[0082] Among these, electricity price fluctuations can be the time-of-use or real-time electricity price changes of the external power grid, serving as an economically sensitive factor to guide the attention mechanism to focus on high-cost periods. Time-series importance weighting can be a mechanism that assigns differentiated weights to historical or future states based on the current scheduling objective, enabling the model to adaptively identify the key time steps with the greatest impact on decision-making. The second scheduling feature can be a context-aware feature generated by the attention mechanism network combining electricity price fluctuations and time-series importance weighting, reflecting the combined impact of external market signals and time sensitivity, thus improving strategy economics. In an exemplary embodiment, net energy demand is input into the attention mechanism network, and the second scheduling feature is extracted by combining electricity price fluctuations and time-series importance weighting. This can be achieved by using net energy demand as the input sequence, combining electricity price signals as the external context, calculating the importance weights of each time step, and then weighting and aggregating them. Furthermore, this operation can be achieved by constructing a query from the electricity price embedding vector and a key / value from the demand sequence, and generating a weighted output through scaling dot product attention; or by using a temporal convolution + softmax weight allocation mechanism to explicitly model the impact of electricity price fluctuations on historical demand states, thereby dynamically highlighting the impact of high electricity prices or key disturbance periods and generating economically sensitive features.

[0083] The first and second scheduling features are merged to generate a comprehensive scheduling feature. Based on this comprehensive scheduling feature, the optimal energy allocation strategy that satisfies economic, stability and carbon emission constraints is solved, and the control commands for each energy unit are output.

[0084] The first and second scheduling features can be two complementary intermediate representations of scheduling, respectively derived from dense connection networks and attention mechanism networks. These can be used as inputs to the fusion operation, providing both structural and contextual perspectives. The comprehensive scheduling feature can be a unified scheduling representation formed by fusing structural nonlinear features and temporal context-weighted features. This can provide a complete input basis for multi-objective optimization solutions. In a specific embodiment, the comprehensive scheduling feature can include, but is not limited to, one or more of spliced ​​fusion features, weighted summation features, and gated fusion features. Economic constraints can be economic objectives such as minimizing energy costs or budget restrictions that the scheduling scheme must meet. These can be used to ensure that the output strategy is financially feasible and cost-effective. Stability constraints can be physical or control boundary conditions that ensure the safe operation of the energy system, such as the number of equipment start-ups and shutdowns, and power ramp-up rates. These can be used to prevent scheduling commands from causing equipment overload or system instability. Carbon emission constraints can be upper limits on the total amount or intensity of greenhouse gas emissions during the scheduling process. These can be used to promote low-carbon operation and support corporate sustainable development goals. For example, carbon emission constraints can include limits on carbon intensity per unit of output, daily cumulative carbon quotas, and upper limits on the proportion of fossil energy use. The control commands for each energy unit can be executable operation signals generated for specific energy equipment (such as boilers, energy storage, photovoltaic inverters, etc.). They can be used to transform abstract scheduling strategies into actions that can be executed by the physical system, thereby achieving closed-loop control.

[0085] The first and second scheduling features are fused to generate a comprehensive scheduling feature. This can be achieved by concatenating the two types of features into vectors, weighted summation, or gated fusion to form a unified representation. Furthermore, this operation can be implemented by concatenating along the feature dimensions and then reducing the dimensionality through a fully connected layer to learn cross-feature interactions; or by using a gating mechanism (such as sigmoid weights) to adaptively adjust the contribution ratio of the two types of features. This integrates structural and contextual information, improving the completeness and discriminative power of the scheduling representation.

[0086] Solving for the optimal energy allocation strategy that satisfies economic, stability, and carbon emission constraints based on this comprehensive scheduling feature can be achieved by inputting the comprehensive scheduling feature into a constrained optimization module (such as a differentiable programming layer or a reinforcement learning policy network), and outputting a policy that satisfies multiple objectives. In a specific embodiment, this operation can be achieved by embedding the constraints into the loss function using the Lagrange relaxation method and solving for the approximate optimal solution through gradient descent; or by integrating a mixed-integer programming solver as a differentiable module and training the policy network parameters end-to-end, enabling collaborative optimization of cost, safety, and environmental objectives within a single framework to generate a feasible and efficient scheduling scheme.

[0087] The output control commands for each energy unit can be used to resolve the optimal energy allocation strategy into operation signals such as start / stop, power setting, and charging / discharging rate of specific equipment. This allows for the mapping from abstract decision-making to physical execution, achieving closed-loop control.

[0088] Taking the day-ahead dispatch of a high-energy-consuming manufacturing enterprise as an example, the dynamic balance management method for enterprise energy supply and demand in this embodiment can be applied to an electrolytic aluminum plant: receiving the net electricity demand forecast for the next 24 hours (including photovoltaic output deduction). First, the net demand sequence is input into a densely connected network to extract the first dispatch feature reflecting load ramping characteristics and equipment inertia; simultaneously, the same sequence is input into an attention mechanism network, and combined with the time-of-use electricity price published on the day (peak-valley price difference of up to 3 times), the high-price period from 18:00 to 22:00 is automatically assigned a higher weight to generate the second dispatch feature. After the two are integrated into a comprehensive dispatch feature, the dispatch model simultaneously minimizes the cost of electricity purchase, limits the number of electrolytic cell start-ups and shutdowns (stability constraints), and ensures that the proportion of purchased coal power does not exceed 30% (carbon emission constraints) during the solution process. Finally, it outputs the rectifier unit power curve, energy storage charging and discharging plan, and standby gas boiler start-up and shutdown instructions, achieving the dual goals of cost reduction and carbon emission compliance.

[0089] This embodiment provides a method for dynamic balance management of enterprise energy supply and demand. It extracts a first scheduling feature by inputting net energy demand into a dense network for nonlinear feature extraction. Then, it inputs net energy demand into an attention mechanism network, weighted by electricity price fluctuations and time-series importance, to extract a second scheduling feature. The first and second scheduling features are then fused to generate a comprehensive scheduling feature. Based on this comprehensive scheduling feature, the optimal energy allocation strategy that satisfies economic, stability, and carbon emission constraints is solved, and control commands for each energy unit are output. The dense network preserves the multi-scale nonlinear structure of energy demand to generate a first scheduling feature sensitive to the system's physical response. The attention mechanism network, weighted by electricity price fluctuations and time-series importance, dynamically focuses on high-cost or critical disturbance periods to generate an economically aware second scheduling feature. By fusing these two complementary features to form a comprehensive scheduling feature with complete information, it serves as the input basis for multi-objective optimization. During the solution process, it simultaneously satisfies the triple constraints of economic, stability, and carbon emission to generate an optimal energy allocation strategy that balances cost, safety, and low carbon emissions. This achieves the technical effect of effectively reducing energy costs and supporting green and low-carbon transformation while ensuring safe system operation.

[0090] In one embodiment, key features are extracted from the fused data based on a multimodal variational autoencoder to obtain key features of the multi-source heterogeneous data, including:

[0091] The encoder maps the fused multi-source heterogeneous data into the probability distribution of latent variables, which represent the implicit operating state of the energy system.

[0092] The fused multi-source heterogeneous data can be a set of multi-energy and external disturbance data that has been processed by a spatiotemporal alignment mechanism and structured and integrated under a unified spatiotemporal benchmark. This data can be used as input to a multimodal variational autoencoder (MAE), providing a foundation for multi-dimensional collaborative modeling. The encoder can be a neural network module in a MAE that maps input data to latent variable distribution parameters (such as mean and variance), enabling nonlinear compression mapping from a high-dimensional observation space to a low-dimensional probabilistic latent space. In an exemplary embodiment, the encoder can be one or more of, but not limited to, modality-specific encoders, shared latent encoders, and graph-aware encoders. Latent variables can be random variables representing the implicit operating state of an energy system in the latent space, following a probability distribution defined by the encoder output parameters. These variables can be used as a compact, decoupled representation of the multi-energy coupling dynamics of the system for feature extraction and reconstruction. The implicit operating state of the energy system can be the internal energy flow and equipment coordination state that cannot be directly observed but can be inferred from multi-source data. This can be used to reflect the abstract semantics of real physical processes, improving the model's interpretability and generalization ability.

[0093] The encoder maps fused multi-source heterogeneous data to the probability distribution of latent variables. This can be achieved by inputting the fused data into the encoder network and outputting the mean and variance parameters of the latent variables, defining their Gaussian distribution. Furthermore, this operation can be implemented through an end-to-end training framework, thereby enabling nonlinear embedding of multimodal data into a structured probabilistic latent space.

[0094] Based on the decoder, the original multimodal data is reconstructed from the distribution of latent variables to form a reconstructed data matrix;

[0095] The decoder can be a generative network in a multimodal variational autoencoder that samples and reconstructs the original multimodal data from latent variables. It can be used to verify the information integrity of the latent representation through inverse mapping and drive feature learning. In a specific embodiment, the decoder can include, but is not limited to, a cross-modal decoder, a conditional generative decoder, and a multi-head output decoder. The reconstructed data matrix can be a reconstructed data tensor output by the decoder with the same dimensions as the original multimodal input. It can be used to calculate the reconstruction error and measure the degree to which the latent representation retains the original information.

[0096] Reconstructing original multimodal data from the latent variable distribution using a decoder can be achieved by sampling from the latent variable distribution and inputting the data into the decoder to generate reconstructed outputs for each modality. For example, this operation can be accomplished by designing a dedicated decoder head for each modality to independently reconstruct data such as electricity consumption, gas consumption, and weather, or by using a unified decoder in conjunction with modality identifier embedding to achieve cross-modal joint reconstruction. This allows for verification of the completeness of the latent representation and supports end-to-end training. Forming the reconstructed data matrix can be achieved by concatenating the reconstruction results of each modality output by the decoder into a tensor isomorphic to the input, thus providing a comparable benchmark format for calculating reconstruction errors.

[0097] By minimizing the reconstruction error and KL divergence loss, a multimodal variational autoencoder is trained, and key features are extracted from the latent representation of the encoder output.

[0098] The reconstruction error can be a measure of the difference between the original fused data and the reconstructed data matrix, typically using mean squared error or cross-entropy. It can be used to constrain model fidelity and prevent excessive information compression that could lead to the loss of key features. The KL divergence loss is the probabilistic distance between the latent variable distribution output by the encoder and a preset prior distribution (such as a standard normal distribution). It can be used to regularize the latent space structure, improving generation capability and feature stability. The latent representation can be a deterministic low-dimensional vector obtained by sampling or taking the expected value from the latent variable distribution. It can be used as a direct source of key features for use by downstream prediction models.

[0099] Training a multimodal variational autoencoder by minimizing reconstruction error and KL divergence loss can jointly optimize reconstruction fidelity and latent space regularity, using stochastic gradient descent to update network parameters. Further, this operation can employ a β-VAE strategy to weight the KL divergence term, controlling the degree of feature decoupling, or introduce adversarial training to assist the decoder and improve reconstruction quality, thereby learning a latent representation that combines information preservation and structural regularity. Extracting key features from the latent representation output by the encoder can be achieved by fixing the trained encoder during the inference phase and directly using its output latent representation as key features. In a specific embodiment, this operation can take the mean vector of the latent distribution as deterministic features, or take statistical aggregations (such as the mean or maximum response) after multiple samplings as robust features, thereby obtaining low-dimensional, denoised, and multimodal semantically integrated key features for use by the prediction model.

[0100] Taking the feature learning of a multi-energy complementary park as an example, the enterprise energy supply and demand dynamic balance management method in this embodiment can be as follows: In a certain integrated energy park, the fused data includes 15-minute granularity data such as power load, natural gas flow, steam pressure, photovoltaic output, outdoor temperature, and production shift labels. The encoder of the multimodal variational autoencoder maps the 7-dimensional time series tensor into a Gaussian distribution of 64-dimensional latent variables, and its mean vector represents the current implicit operating state of "high heat production - low photovoltaic power". The decoder reconstructs the data of each modality to form a reconstruction matrix. During the training process, the MSE reconstruction error and KL divergence of each modality are jointly minimized. Finally, the latent mean is taken as the key feature input to the prediction model, which is significantly better than the baseline method that directly uses the original fused data.

[0101] This embodiment provides a method for dynamic balance management of enterprise energy supply and demand. It maps fused multi-source heterogeneous data into a probability distribution of latent variables based on an encoder, reconstructs the original multimodal data from the latent variable distribution based on a decoder to form a reconstructed data matrix, and trains a multimodal variational autoencoder by minimizing reconstruction error and KL divergence loss and extracts key features from the latent representation output by the encoder. By compressing high-dimensional heterogeneous observations into a structured probabilistic latent representation, verifying information integrity using a reconstruction task, and jointly optimizing fidelity and latent space regularity to obtain stable low-dimensional features, this method can achieve the technical effects of effectively fusing multimodal energy and disturbance data, extracting denoised and semantically rich key features, enhancing the ability to model the implicit operating state of the system, and improving the accuracy of subsequent prediction and scheduling decisions.

[0102] In one embodiment, the training process of the energy supply and demand balance prediction model includes:

[0103] Based on the stochastic gradient Hamiltonian Monte Carlo sampling method, the weight parameters of the initial energy supply and demand balance prediction model are sampled under the Bayesian framework to achieve posterior distribution approximation, complete the preliminary parameter adjustment, and obtain the energy supply and demand balance prediction model after preliminary parameter adjustment.

[0104] Furthermore, stochastic gradient Langevin dynamics is introduced to inject controllable noise during parameter updates, enhancing the model's generalization ability and avoiding getting trapped in local optima. The model parameters are adaptively optimized, ultimately resulting in a robust energy supply and demand balance prediction model.

[0105] The initial energy supply and demand balance prediction model can be an original prediction model structure consisting of a deep kernel learning neural network improved by kernel function and a deep Gaussian process that has not yet been trained by Bayes. It can be used as the starting point for Bayesian parameter sampling and optimization, providing a trainable weight parameter space.

[0106] Weight parameters can be a set of parameters consisting of all learnable connection weights and bias terms in an energy supply and demand balance prediction model. They can be used to determine the mapping relationship from model input to output, and their distribution characteristics directly affect prediction performance and uncertainty expression.

[0107] The Bayesian framework is a probabilistic modeling paradigm that treats model parameters as random variables and derives the posterior distribution based on prior and likelihood. It can be used to provide uncertainty quantification capabilities for prediction models and avoid overfitting and lack of confidence caused by point estimation.

[0108] Stochastic gradient Hamiltonian Monte Carlo (SMR) sampling methods can be Markov chain Monte Carlo sampling algorithms that combine Hamiltonian dynamics and stochastic gradient estimation for efficient exploration of the posterior distribution of high-dimensional parameters. In this embodiment, SMR sampling methods can simulate the evolution of the physical system by introducing momentum variables, performing non-random walk sampling under gradient guidance to reduce sample correlation. For example, SMR sampling methods may include, but are not limited to, one or more of No-U-TurnSampler (NUTS), SMR variational inference, and hierarchical Hamiltonian Monte Carlo methods.

[0109] The posterior distribution approximation can be a numerical approximation of the true posterior probability distribution of model parameters under given observation data conditions in Bayesian inference. It can be used to replace the true posterior, which is difficult to calculate analytically, and to generate predictive outputs with uncertainty quantification.

[0110] The energy supply and demand balance prediction model after preliminary parameter adjustment can be an intermediate model state after the first round of Bayesian parameter updates is completed through stochastic gradient Hamiltonian Monte Carlo sampling. It can be used to retain parameter uncertainty information and provide an initial basis for subsequent optimization by introducing dynamic noise.

[0111] Based on the stochastic gradient Hamiltonian Monte Carlo (SGR) sampling method, the weight parameters of the initial energy supply and demand balance prediction model are sampled within a Bayesian framework. This can be achieved by using the SGR algorithm to extract multiple samples from the posterior distribution of the weight parameters under Bayesian constraints. Furthermore, this operation can be implemented by estimating the gradient using mini-batch data and simulating the Hamiltonian trajectory using momentum variables to generate low-correlation parameter samples, or by using a No-U-Turn mechanism to automatically adjust the trajectory length to avoid manually setting the number of sampling steps. This allows for obtaining multiple sets of weight configurations that reflect parameter uncertainty, supporting subsequent probabilistic predictions.

[0112] To achieve posterior distribution approximation and complete preliminary parameter adjustments, the weighted sample set obtained from sampling can be used as a posterior approximation to initialize or replace the original model parameters, enabling the model to have a preliminary ability to perceive uncertainty and get rid of the limitation of a single optimal solution.

[0113] The energy supply and demand balance prediction model obtained after preliminary parameter adjustment can be applied to the original model structure by integrating the posterior approximate corresponding parameters (such as average or integrated prediction) to form an intermediate prediction model with Bayesian properties, which provides a basis for the second stage of optimization.

[0114] Controllable noise can be a random perturbation signal injected during parameter update according to a preset variance or adaptive strategy. It can be used to break the deterministic optimization path, help the model escape the local optimum region, and improve generalization ability.

[0115] Stochastic gradient Langevin dynamics can be achieved by incorporating Gaussian noise conforming to the Langevin equation into gradient descent updates to approximate the target distribution. In an exemplary embodiment, the parameter update method of stochastic gradient Langevin dynamics can be manifested as follows: the current parameter is updated by superimposing a deterministic term proportional to the logarithmic posterior gradient and a random Gaussian perturbation term proportional to the square root of the step size. Exemplarily, stochastic gradient Langevin dynamics can include, but is not limited to, one or more of the following: annealed Langevin dynamics, spectral normalized Langevin sampling, and adaptive step-size Langevin dynamics.

[0116] A robust energy supply and demand balance forecasting model can be a final forecasting model that has strong anti-disturbance ability and stable forecasting performance after two-stage Bayesian optimization. It can be used to output highly reliable net energy demand forecasts even under uncertain environments such as renewable energy fluctuations and changes in production plans.

[0117] By introducing stochastic gradient Langevin dynamics, controllable noise is injected into the parameter update process. This can be achieved by superimposing Gaussian noise with a specific variance into the gradient-based parameter update step. Furthermore, this operation can be implemented through global exploration with a fixed noise variance, or by dynamically adjusting the noise intensity according to the curvature of the training loss to achieve local adaptive perturbation. This can break the deterministic optimization path and improve the model's ability to escape local optima.

[0118] To enhance the model's generalization ability and avoid getting trapped in local optima, noise perturbation can be used to make the optimization trajectory cover a wider range of parameter space, thereby improving the model's predictive stability under unseen conditions (such as extreme weather or equipment failure).

[0119] Adaptive optimization of model parameters can be achieved by combining gradient direction and noise perturbation, and dynamically adjusting the parameter update step size and direction so that the model can continuously adapt to the time-varying characteristics of the energy system's operating state.

[0120] The final robust energy supply and demand balance prediction model can be solidified after completing two-stage Bayesian optimization or integrated into the final parameter set to form a deployment model, so that the output has high accuracy, high reliability and strong anti-disturbance capability in net energy demand prediction.

[0121] Taking the predictive modeling of manufacturing enterprises with a high proportion of renewable energy access as an example, the dynamic balance management method for enterprise energy supply and demand in this embodiment can be deployed in an electronics manufacturing plant. First, an initial energy supply and demand balance prediction model composed of a deep kernel network and a deep Gaussian process is constructed. Within a Bayesian framework, stochastic gradient Hamiltonian Monte Carlo sampling is used to sample the model weights 500 times to obtain an approximate posterior distribution, forming a model with preliminary parameter adjustments. Then, stochastic gradient Langevin dynamics is introduced, injecting Gaussian noise with a variance of 0.01 at each step in the subsequent 100 training rounds, ensuring the model maintains predictive stability even under scenarios such as sudden drops in photovoltaic output or surges in orders. The resulting robust prediction model improves the coverage of the 24-hour net electricity demand prediction interval from 95% to 92% on the test set, significantly outperforming traditional point estimation models. The reliable prediction results output by this model are directly used in the scheduling optimization module, supporting the accurate formulation of energy storage charging and discharging and external power purchase strategies.

[0122] This embodiment provides a method for dynamic energy supply and demand balance management in enterprises. It samples the weight parameters of an initial energy supply and demand balance prediction model within a Bayesian framework using a stochastic gradient Hamiltonian Monte Carlo sampling method, approximating the posterior distribution and performing preliminary parameter adjustments to obtain a pre-adjusted energy supply and demand balance prediction model. Furthermore, it introduces stochastic gradient Langevin dynamics to inject controllable noise during parameter updates, enhancing the model's generalization ability and avoiding local optima. This adaptively optimizes the model parameters to ultimately obtain a robust energy supply and demand balance prediction model. By efficiently approximating complex posterior distributions using the stochastic gradient Hamiltonian Monte Carlo sampling method to preserve parameter uncertainty and support probabilistic prediction, and by superimposing controllable noise on gradient-guided basis using stochastic gradient Langevin dynamics to maintain the exploration-utilization balance and improve adaptability to multimodal non-stationary energy data, it can significantly improve the model's prediction accuracy, reliability, and disturbance resistance in high-dimensional, nonlinear, and highly uncertain industrial scenarios. This provides a highly reliable net energy demand prediction basis for scheduling decisions, thereby supporting the overall management method to achieve synergistic optimization of economy, energy efficiency, and system safety.

[0123] In one embodiment, the training process of the energy dispatch optimization decision model includes:

[0124] During backpropagation, an adaptive gradient pruning mechanism is used to dynamically adjust the gradient pruning threshold by using the second moment estimation of the gradient to prevent gradient explosion, thus obtaining an energy scheduling optimization decision model with preliminary parameter adjustments.

[0125] Furthermore, an adaptive momentum estimation algorithm is adopted to dynamically adjust the learning rate and momentum factor based on the first and second moments of the gradient, thereby improving the convergence speed and stability of the model and finally training a high-performance energy dispatch optimization decision model.

[0126] The backpropagation process can be an algorithmic flow in neural network training that calculates the gradient of the loss function with respect to the parameters layer by layer from the output layer to the input layer using the chain rule. It can be used to provide a basis for the direction and magnitude of model parameter updates.

[0127] Gradient second-moment estimation can be an online estimation of the expected value of the squared gradient, reflecting the variance characteristics of gradient fluctuations. It can be used as a statistical basis for dynamically setting the clipping threshold in adaptive gradient clipping mechanisms. In an exemplary embodiment, gradient second-moment estimation may include, but is not limited to, one or more of the following: moving average second-moment, exponentially decaying second-moment, and window-truncation second-moment.

[0128] Gradient clipping thresholds can be used to limit the numerical boundary of the upper bound of the gradient norm, preventing excessive parameter updates. They can also be used to control training stability and prevent the model from diverging due to extreme gradients.

[0129] An adaptive gradient pruning mechanism can be a gradient stabilization method that dynamically adjusts the pruning threshold based on gradient statistical properties (such as the second moment). In this embodiment, the adaptive gradient pruning mechanism can calculate the gradient second moment estimate in each round of backpropagation, and set the current pruning threshold accordingly, replacing the fixed threshold. Furthermore, the adaptive gradient pruning mechanism can generate a threshold based on the gradient second moment estimate, apply it to the original gradient during backpropagation, and output the pruned gradient for parameter updates.

[0130] The energy dispatch optimization decision model after initial parameter adjustment can be an intermediate training version model after adaptive gradient pruning, which can be used as the initialization basis for the subsequent adaptive momentum optimization stage.

[0131] During backpropagation, an adaptive gradient clipping mechanism is used to dynamically adjust the gradient clipping threshold using gradient second-moment estimation. This can be achieved by estimating the gradient second moment after each backpropagation calculation, then setting the current clipping threshold accordingly, and performing norm clipping on the gradient. Furthermore, this operation can be implemented by using L2 norm clipping with the threshold set as the square root of the gradient second-moment estimate multiplied by an adjustable coefficient, or by using quantile estimation instead of the second moment and setting the threshold as the dynamic percentile of the gradient magnitude. This allows the clipping intensity to adaptively change with the gradient distribution, avoiding over- or under-clipping.

[0132] Preventing gradient explosion can be achieved by limiting the magnitude of gradient updates, ensuring that parameter changes remain within a reasonable range, thereby guaranteeing numerical stability during training and preventing the loss function from diverging.

[0133] The energy dispatch optimization decision model obtained after preliminary parameter adjustment can be updated once using the adaptively pruned gradient to generate an intermediate model version, thereby obtaining an initial optimization model with stable gradient and complete structure, laying the foundation for subsequent fine-tuning.

[0134] The first moment of the gradient can be the expected estimate of the gradient, usually maintained in the form of an exponential moving average. It can be used to reflect the long-term trend direction of the gradient and for dynamic adjustment of the momentum factor and learning rate. The second moment of the gradient can be the expected estimate of the squared gradient, representing the local curvature or uncertainty in each parameter dimension. It can be used to adaptively scale the learning rate, achieving step size adjustment at the parameter dimension level.

[0135] The learning rate is a hyperparameter that controls the step size of each model parameter update. It can affect the model's convergence speed and final performance and needs to be dynamically adjusted during training. The momentum factor is a coefficient used to accumulate historical gradient directions to accelerate convergence and suppress oscillations. It can be used to improve the smoothness of the optimization path and its ability to traverse flat regions.

[0136] Adaptive momentum estimation algorithms can be optimizer variants that dynamically adjust the learning rate and momentum factor by combining the first and second moments of the gradient. In one specific embodiment, the adaptive momentum estimation algorithm can maintain an exponential moving average of the first and second moments of the gradient (mean) and calculate parameter-specific learning rates and momentum weights accordingly. Exemplary examples include, but are not limited to, bias-corrected adaptive momentum, hierarchical momentum estimation, and constraint-aware momentum adjustment.

[0137] A high-performance energy dispatch optimization decision model can be a final model obtained after a complete adaptive optimization training process, possessing high convergence stability and strategy generation capabilities. It can be used to reliably output the optimal energy allocation strategy that balances economy, energy efficiency, and system safety in actual operation.

[0138] An adaptive momentum estimation algorithm is employed, dynamically adjusting the learning rate and momentum factor based on the first and second moments of the gradient. This can be achieved during the parameter update phase by calculating an independent learning rate for each parameter dimension based on the maintained first and second moment estimates, and dynamically adjusting the momentum weights. Furthermore, this operation can be further refined by borrowing from the AdamW framework, separating momentum updates outside the weight decay term (the momentum factor increases with the consistency of the first moment), or by introducing scheduling-aware constraints to reduce the learning rate and weaken momentum realization in directions that violate physical feasibility. This allows for efficient and stable parameter search on complex energy scheduling target surfaces. Improving model convergence speed and stability can be achieved by avoiding oscillations through adaptive learning rates and optimizing paths through momentum smoothing, thereby shortening training time and improving the reliability of policy generation in multi-energy coupling scenarios. Finally, a high-performance energy scheduling optimization decision model can be obtained by saving the converged model parameters after all training iterations, thus outputting a highly robust deep optimization model suitable for real-time scheduling decisions.

[0139] Taking the training of a scheduling model for a high-fluctuation renewable energy park as an example, the enterprise energy supply and demand dynamic balance management method in this embodiment can be used to train the energy scheduling optimization decision model of a manufacturing park with 50% photovoltaic penetration. During backpropagation, sudden changes in photovoltaic output cause a surge in gradient amplitude during certain periods. Traditional fixed threshold pruning either excessively suppresses effective signals or fails to prevent the surge. This scheme adopts an adaptive gradient pruning mechanism, which calculates the second-order moment estimate of the gradient in real time (such as the variance of the power command gradient of each device), dynamically sets the pruning threshold, and effectively stabilizes the initial training. Subsequently, an adaptive momentum estimation algorithm is introduced. During complex periods where electricity price peaks and photovoltaic drops overlap, the model automatically reduces the learning rate of relevant parameters, enhances momentum consistency, and avoids policy oscillations. The final model converges within 200 rounds, reducing the number of iterations by 40% compared to standard Adam training, and the probability of the scheduling scheme violating the equipment ramp-up rate constraint on the test set decreases by 60%, significantly improving the reliability of actual deployment.

[0140] This embodiment provides a method for dynamic balance management of enterprise energy supply and demand. During backpropagation, an adaptive gradient pruning mechanism is used to dynamically adjust the gradient pruning threshold using second-order moment estimation to prevent gradient explosion, resulting in an energy dispatch optimization decision model with preliminary parameter adjustments. Furthermore, an adaptive momentum estimation algorithm is employed to dynamically adjust the learning rate and momentum factor based on the first and second moments of the gradient, improving the model's convergence speed and stability. Finally, a high-performance energy dispatch optimization decision model is trained. By adaptively scaling the pruning threshold according to the statistical characteristics of the gradient distribution to suppress the destructive effects of extreme gradients, and by fusing first and second-order moment information to achieve adaptive step size and direction smoothing at the parameter dimension level, this method significantly improves training stability, accelerates convergence speed, and enhances the reliability of strategy generation in high-dimensional nonlinear, multi-constraint energy dispatch tasks, thereby supporting the real-time, safe, and economical operation of intelligent energy systems.

[0141] Furthermore, to achieve the above objectives, the present invention also provides an enterprise energy supply and demand dynamic balance management and control system, the system comprising:

[0142] The data acquisition module is used to collect multi-source heterogeneous data from the enterprise. The multi-source heterogeneous data includes one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data.

[0143] The feature fusion module is used to perform spatiotemporal alignment and data fusion on the multi-source heterogeneous data, and to extract key features from the fused data based on a multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data.

[0144] The demand forecasting module is used to input the key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance forecasting model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance forecasting model. The energy supply and demand balance forecasting model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process.

[0145] The scheduling decision module is used to input the net energy demand into a pre-built energy scheduling optimization decision model to obtain the optimal energy allocation strategy output by the energy scheduling optimization decision model. The energy scheduling optimization decision model is constructed based on a dense connection network combined with an attention mechanism network.

[0146] Other embodiments or specific implementations of the enterprise energy supply and demand dynamic balance management and control system described in this invention can be referred to the above-mentioned method embodiments, and will not be repeated here.

[0147] In addition, to achieve the above objectives, the present invention also provides an enterprise energy supply and demand dynamic balance management and control device, the device comprising: a memory, a processor, and an enterprise energy supply and demand dynamic balance management and control program stored in the memory and executable on the processor, the enterprise energy supply and demand dynamic balance management and control program being configured to implement the steps of the enterprise energy supply and demand dynamic balance management and control method as described above.

[0148] In addition, to achieve the above objectives, the present invention also provides a medium storing an enterprise energy supply and demand dynamic balance management program, wherein when the enterprise energy supply and demand dynamic balance management program is executed by a processor, it implements the steps of the enterprise energy supply and demand dynamic balance management method as described above.

[0149] In addition, to achieve the above objectives, the present invention also provides a computer program product, including an enterprise energy supply and demand dynamic balance management and control program, wherein when the enterprise energy supply and demand dynamic balance management and control program is executed by a processor, it implements the steps of the enterprise energy supply and demand dynamic balance management and control method as described above.

[0150] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for dynamic balance management and control of energy supply and demand in enterprises, characterized in that, The method includes: Collect multi-source heterogeneous data from enterprises, including one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data. The multi-source heterogeneous data is spatiotemporally aligned and fused, and key features are extracted from the fused data based on a multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data. The key features and the enterprise's physical topology parameters are input into a pre-built energy supply and demand balance prediction model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance prediction model. The energy supply and demand balance prediction model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process. The net energy demand is input into a pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model. The energy dispatch optimization decision model is built based on a dense connection network combined with an attention mechanism network.

2. The enterprise energy supply and demand dynamic balance management method as described in claim 1, characterized in that, The process of inputting the key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance prediction model to obtain the enterprise's net energy demand for future periods output by the energy supply and demand balance prediction model includes: The key features and the enterprise's physical topology parameters are input into a deep kernel learning neural network for deep feature extraction to obtain the deep features of the key features. Based on the kernel function, the deep features are mapped to a high-dimensional feature space to obtain high-dimensional deep features; Net energy demand is predicted based on the high-dimensional deep features using a deep Gaussian process, yielding the enterprise's net energy demand over a predetermined future time period.

3. The enterprise energy supply and demand dynamic balance management method as described in claim 1, characterized in that, The step of inputting the net energy demand into a pre-built energy dispatch optimization decision model to obtain the optimal energy allocation strategy output by the energy dispatch optimization decision model includes: The net energy demand is input into a dense network for nonlinear feature extraction to obtain the first scheduling feature. The net energy demand is input into the attention mechanism network, and the second scheduling feature is extracted by weighting the electricity price fluctuation and the temporal importance. The first and second scheduling features are merged to generate a comprehensive scheduling feature. Based on this comprehensive scheduling feature, the optimal energy allocation strategy that satisfies economic, stability and carbon emission constraints is solved, and control commands for each energy unit are output.

4. The enterprise energy supply and demand dynamic balance management method as described in claim 1, characterized in that, The key features of the multi-source heterogeneous data are obtained by extracting key features from the fused data based on the multimodal variational autoencoder, including: The encoder maps the fused multi-source heterogeneous data into a probability distribution of latent variables, which represent the implicit operating state of the energy system. Based on the decoder, the original multimodal data is reconstructed from the distribution of the latent variables to form a reconstructed data matrix; By minimizing the reconstruction error and KL divergence loss, a multimodal variational autoencoder is trained, and key features are extracted from the latent representation of the encoder output.

5. The enterprise energy supply and demand dynamic balance management method as described in claim 1, characterized in that, The training process of the energy supply and demand balance prediction model includes: Based on the stochastic gradient Hamiltonian Monte Carlo sampling method, the weight parameters of the initial energy supply and demand balance prediction model are sampled under the Bayesian framework to achieve posterior distribution approximation, complete the preliminary parameter adjustment, and obtain the energy supply and demand balance prediction model after preliminary parameter adjustment. Furthermore, stochastic gradient Langevin dynamics is introduced to inject controllable noise during parameter updates, enhancing the model's generalization ability and avoiding getting trapped in local optima. The model parameters are adaptively optimized, ultimately resulting in a robust energy supply and demand balance prediction model.

6. The enterprise energy supply and demand dynamic balance management method as described in claim 1, characterized in that, The training process of the energy dispatch optimization decision model includes: During backpropagation, an adaptive gradient pruning mechanism is used to dynamically adjust the gradient pruning threshold by using the second moment estimation of the gradient to prevent gradient explosion, thus obtaining an energy scheduling optimization decision model with preliminary parameter adjustments. Furthermore, an adaptive momentum estimation algorithm is adopted to dynamically adjust the learning rate and momentum factor based on the first and second moments of the gradient, thereby improving the convergence speed and stability of the model and finally training a high-performance energy dispatch optimization decision model.

7. A dynamic balance management and control system for enterprise energy supply and demand, characterized in that, The system includes: The data acquisition module is used to collect multi-source heterogeneous data from the enterprise. The multi-source heterogeneous data includes one or more of the following: internal electricity load data, gas consumption data, heat consumption data, production equipment operating status data, energy storage system status data, renewable energy power generation data, external power grid electricity price signals, meteorological and environmental data, and enterprise production plan data. The feature fusion module is used to perform spatiotemporal alignment and data fusion on the multi-source heterogeneous data, and to extract key features from the fused data based on a multimodal variational autoencoder to obtain the key features of the multi-source heterogeneous data. The demand forecasting module is used to input the key features and the enterprise's physical topology parameters into a pre-built energy supply and demand balance forecasting model to obtain the net energy demand of the enterprise in the future period output by the energy supply and demand balance forecasting model. The energy supply and demand balance forecasting model is built based on a deep kernel learning neural network with kernel function improvement combined with a deep Gaussian process. The scheduling decision module is used to input the net energy demand into a pre-built energy scheduling optimization decision model to obtain the optimal energy allocation strategy output by the energy scheduling optimization decision model. The energy scheduling optimization decision model is constructed based on a dense connection network combined with an attention mechanism network.

8. A dynamic energy supply and demand balance management and control device for enterprises, characterized in that, The device includes: a memory, a processor, and an enterprise energy supply and demand dynamic balance management program stored in the memory and executable on the processor, the enterprise energy supply and demand dynamic balance management program being configured to implement the steps of the enterprise energy supply and demand dynamic balance management method as described in any one of claims 1 to 6.

9. A medium, characterized in that, The medium stores an enterprise energy supply and demand dynamic balance management program, which, when executed by a processor, implements the steps of the enterprise energy supply and demand dynamic balance management method as described in any one of claims 1 to 6.

10. A computer program product, comprising an enterprise energy supply and demand dynamic balance management and control program, characterized in that, When the enterprise energy supply and demand dynamic balance management program is executed by the processor, it implements the steps of the enterprise energy supply and demand dynamic balance management method as described in any one of claims 1 to 6.

Images