CN-122022297-A - Multi-energy collaborative scheduling optimization method and system based on deep learning

CN122022297ACN 122022297 ACN122022297 ACN 122022297ACN-122022297-A

Abstract

The invention discloses a multi-energy collaborative scheduling optimization method and a system based on deep learning in the technical field of collaborative management, wherein the method comprises the following steps of S1, acquiring multi-source operation data of each functional unit in a campus organization, and analyzing and extracting decision support information; the method comprises the steps of S2, constructing and training a scheduling decision model, embedding system operation rules, adopting a prediction and decision integrated joint optimization mechanism, S3, training the model based on a meta-learning framework to generate initial parameters, and carrying out quick fine adjustment on the model by using a small amount of new data when the working condition of the system changes, and S4, carrying out performance optimization and deployment on the model to generate a real-time energy scheduling plan. According to the method, decision security is ensured through an endophytic embedded system operation rule, scheduling robustness and optimality are improved through prediction-decision integrated optimization, dynamic adaptability of a model is enhanced through a meta-learning mechanism, and operation efficiency, security and long-term stability of a multi-energy system are integrally improved.

Inventors

ZHANG TAO
WANG LIANG
XIAO JUNBO
CHEN ZHONG
GU SHAOCHEN
LIU JIAPENG
ZHANG MING

Assignees

淮阴工学院

Dates

Publication Date: 20260512
Application Date: 20260121

Claims (10)

1. The multi-energy collaborative scheduling optimization method based on deep learning is characterized by comprising the following steps of: S1, acquiring multi-source operation data of each functional unit in a campus organization, analyzing the multi-source operation data, and extracting decision support information representing dynamic association relation between the functional units and the organization; S2, constructing and training an end-to-end scheduling decision model based on the decision support information, explicitly encoding a preset system operation rule into a scheduling decision model structure, and simultaneously adopting a prediction and decision integrated joint optimization mechanism to cooperatively process a decision main task and a prediction auxiliary task; S3, training the scheduling decision model based on a meta learning framework to generate initial parameters, and adjusting the scheduling decision model by using new data when the working condition of the system changes; And S4, performing performance optimization and deployment on the scheduling decision model to generate a real-time energy scheduling plan.
2. The method for optimizing multi-energy collaborative scheduling according to claim 1, wherein the step S1 further comprises a step of preprocessing data after the multi-source operation data is obtained, comprising: Smoothing the data by using a moving average filter to filter out high-frequency noise; Detecting abnormal values by adopting a3 sigma principle, and filling the abnormal values and the missing values by adopting a linear interpolation method; all numerical data is scaled into dimensionless intervals by a min-max normalization method.
3. The multi-energy collaborative scheduling optimization method according to claim 2, wherein in step S1, the multi-source operation data is analyzed, specifically: In the space dimension, constructing an association network structure representing the mutual association of each functional unit and the organization of the campus, processing the association network structure through a graph attention mechanism, and identifying the inherent association dependency relationship; in the time dimension, processing the data time sequence of each functional unit and the organization through an attention mechanism, and identifying a dynamic evolution rule of the time sequence; and fusing the association dependency relationship of the space dimension with the dynamic evolution rule of the time dimension to form decision support information containing space-time characteristics.
4. The multi-energy collaborative scheduling optimization method according to claim 1, wherein in step S2, a preset system operation rule is explicitly encoded into the scheduling decision model structure, specifically: Embedding a constraint checking module behind an original output layer of the scheduling decision model, wherein the constraint checking module is embedded with system operation rules; The constraint checking module takes an energy management instruction output by the model as input, and calculates the quantized deviation degree of the energy management instruction and the system operation rule; and adding the quantized deviation degree as a continuous and differentiable penalty term into a total loss function of model training, and guiding model parameters to optimize in a direction meeting constraints through a back propagation algorithm.
5. The method for collaborative scheduling optimization according to claim 4, wherein the prediction and decision-making integrated joint optimization mechanism in step S2 specifically includes: constructing a unified objective function of weighted summation of the main task loss and the auxiliary task loss, evaluating the operation efficiency of a scheduling scheme by the main task loss, and evaluating the uncertainty variable prediction precision by the auxiliary task loss; The scheduling decision main task and the uncertainty prediction auxiliary task share a bottom layer space-time feature extraction network; The model parameters are optimized through end-to-end gradient descent, so that the space-time feature extraction network learns to feature representation which is beneficial to improving decision-making efficiency and prediction accuracy.
6. The multi-energy collaborative scheduling optimization method according to claim 1, wherein the training of the scheduling decision model based on a meta-learning framework in step S3 is specifically: a scene task set is constructed, wherein key indexes are extracted from historical operation data, typical operation modes are partitioned through K-Means clustering, and the scene task set comprising various campus operation modes is constructed; And the outer layer circulation calculates element loss according to the performance loss of all temporary models, and updates the initial parameters to generate model parameters with generalization capability.
7. The method for optimizing multi-energy collaborative scheduling according to claim 6, wherein the specific step of performing the fast fine tuning on the scheduling decision model in the step S3 when the system working condition is changed is as follows: Deploying a working condition change monitor, and judging whether the working condition changes or not by calculating the deviation between the real-time operation data and the historical baseline data; when the working condition change is judged, loading initial parameters generated by meta training, and freezing network parameters of the bottom layer and the middle layer of the model; And performing gradient adjustment on the model top network parameters by using a small amount of new working condition data to generate a scheduling decision model adapting to the new working condition.
8. The multi-energy collaborative scheduling optimization system based on deep learning is characterized by adopting the multi-energy collaborative scheduling optimization method according to any one of claims 1-7, and further comprising: The operation data processing and analyzing module is responsible for the collection, pretreatment and deep analysis of multi-source operation data and extracts decision support information containing space-time characteristics; The resource management optimization module is used for constructing an end-to-end scheduling decision model based on the decision support information and realizing the integrated joint optimization of the endogenous embedding and the predictive decision of the system operation rules; The dynamic strategy self-adaptation module trains the model through the meta-learning framework and utilizes a small amount of new data to quickly fine-tune the model when the working condition of the system changes; And the scheduling plan deployment module is used for performing performance optimization and deployment on the model, generating a real-time energy scheduling plan meeting the requirement of multiple service periods and issuing and executing.
9. The multi-energy collaborative scheduling optimization system according to claim 8, wherein the operation data processing and analysis module specifically comprises: the data acquisition unit integrates various sensors and external API interfaces and acquires resource consumption, power generation data, equipment states, weather and cost signals; the data preprocessing unit is used for smoothing, outlier processing and normalization processing of the original data; the space-time feature extraction unit is used for extracting high-dimensional features integrating space service dependence and time evolution rules through a graph neural network and a time sequence model based on an attention mechanism; The resource management optimization module specifically comprises: the model construction unit is used for constructing an end-to-end scheduling decision model based on a deep reinforcement learning algorithm; the constraint embedding unit compiles the system safe operation rule into a differentiable constraint layer and integrates the differentiable constraint layer at the output end of the model; And the joint optimization unit is used for constructing a unified objective function and realizing joint training and optimization of scheduling decision and uncertainty prediction.
10. The multi-energy collaborative scheduling optimization system according to claim 8, wherein the dynamic policy adaptation module specifically comprises: The meta learning training unit is used for constructing a scene task set and generating initial parameters with generalization capability through a double-layer optimization process training model; the working condition monitoring unit is used for monitoring the running state of the system in real time and judging whether the working condition changes or not; The quick fine tuning unit freezes the general knowledge parameters of the model when the working condition changes and quickly adjusts the top layer parameters by using a small amount of new data; the dispatch plan deployment module specifically includes: The model optimizing unit optimizes the model performance through a model pruning and quantization technology; generating a day-ahead optimal scheduling plan and performing intra-day rolling correction; and the instruction issuing unit issues the scheduling instruction to the terminal equipment controller through the industrial Internet of things gateway.

Description

Multi-energy collaborative scheduling optimization method and system based on deep learning Technical Field The invention relates to the technical field of multi-energy system collaborative management optimization, in particular to a multi-energy collaborative scheduling optimization method and system based on deep learning. Background The realization of the collaborative scheduling optimization of the multiple energy sources is especially the key for improving the operation efficiency and the service quality of the intelligent campus and the like in complex systems comprising various functional organizations such as teaching buildings, dormitories, laboratories and the like and provided with heterogeneous functional units such as distributed photovoltaic, energy storage, charging piles and the like. The prior art has remarkable progress in improving the intelligent level of dispatching by combining a deep learning model with the control of an energy system, and can autonomously learn a dispatching strategy from massive historical data based on a Deep Reinforcement Learning (DRL) method, and a two-stage framework based on prediction-optimization can utilize the strong time sequence prediction capability of deep learning to provide decision basis for a traditional mathematical planning model, thereby laying a solid technical foundation for solving the complex dispatching problem by utilizing a data driving idea by comprehensively applying the technologies. Meanwhile, in a prediction-optimization framework, the inherent decoupling between the prediction model and the optimization model leads to unidirectional propagation and accumulation of prediction errors to a decision end, the information processing flow lacks an effective feedback mechanism from an optimization result to the prediction model, and when severe fluctuation or load mutation occurs to energy sources, the scheduling plan based on the splitting model is difficult to ensure the robustness and the optimality of the final decision. In summary, although the prior art provides a data-driven solution framework for multi-energy collaborative scheduling, there are significant drawbacks in the internalization processing of high-dimensional complex constraints, end-to-end integration fusion of prediction and decision-making, and rapid adaptive capacity to dynamic uncertain environments, which cannot meet the requirements of modern energy systems on security, feasibility and robustness for refinement and dynamics. Disclosure of Invention The multi-energy collaborative scheduling optimization method and system based on deep learning provided by the application solve the defects of the prior art in terms of decision security, global optimality and dynamic adaptability, and improve the running efficiency, security and robustness of a multi-energy system. The embodiment of the application provides a multi-energy collaborative scheduling optimization method based on deep learning, which comprises the following steps: S1, acquiring multi-source operation data of each functional unit in a campus organization, analyzing the multi-source operation data, and extracting decision support information representing dynamic association relation between the functional units and the organization; S2, constructing and training an end-to-end scheduling decision model based on the decision support information, explicitly encoding a preset system operation rule into a scheduling decision model structure, enabling model endogenously learning to meet the system operation rule in the training process, and simultaneously adopting a prediction and decision integrated joint optimization mechanism to cooperatively process a resource scheduling decision main task and a key uncertainty prediction auxiliary task; S3, training the scheduling decision model based on a meta learning framework to generate initial parameters with generalization capability, and performing quick fine adjustment on the scheduling decision model by using a small amount of new data when the working condition of the system changes; And S4, performing performance optimization on the scheduling decision model, and deploying the scheduling decision model in an energy management information system to generate a real-time energy scheduling plan meeting the requirement of multiple service periods. The beneficial effects of the above embodiment are that: The method solves the core defects of the prior art in the aspects of complex constraint processing, prediction and decision decoupling and dynamic adaptability. The model is continuously perceived by a model structure through embedding a system operation rule in the model structure, so that low-efficiency exploration caused by an external sparse punishment signal is avoided, the safety and feasibility of decision making are ensured, the barrier of unidirectional propagation of a prediction error in a traditional two-stage method is broken through prediction and decision