Energy management and control system and method for integrating electric vehicles and renewable energy

Abstract

The present invention relates to the field of control system technology, and specifically discloses an energy management and control system and method for integrating electric vehicles and renewable energy, including an acquisition module, an optimization control module, a control execution module, a constraint module, and a fault-tolerant module. The acquisition module collects photovoltaic, wind power, and electric vehicle load data through a distributed sensor network, and the control execution module issues instructions and provides closed-loop feedback through a multi-priority message queue; the constraint module decouples spatiotemporal constraints through causal reasoning and combines federated learning to update strategy network parameters; the fault-tolerant module is based on dynamic fault tolerance based on Bayesian networks and generative adversarial networks. The method covers multi-source data processing, hierarchical optimization instruction generation, closed-loop verification, and anti-interference training processes, solving the difficulties in solving high-dimensional mixed integer programming and the problem of optimization path deviation caused by spatiotemporal correlation interference, thereby achieving dynamic balance of energy supply and demand and low-carbon operation.

Description

Technical Field

[0001] The present invention relates to the technical field of control systems, and in particular to an energy management and control system and method for integrating electric vehicles and renewable energy. Background Art

[0002] An energy management and control system integrating electric vehicles and renewable energy utilizes multi-level collaborative optimization technology to achieve efficient control of distributed energy resources, energy storage units, and flexible loads. Based on a real-time data acquisition module and a wide-area communication network, the system dynamically captures the output characteristics of intermittent power sources such as photovoltaic and wind power, as well as the load demands of the electric vehicle cluster. Using stochastic optimization theory, a mixed integer programming model with time-coupling constraints is established. A model predictive control (MPC) algorithm is used to roll-correct renewable energy forecast errors. Incorporating the rapid power response characteristics of the electric vehicle's bidirectional converter, a multi-objective optimization function incorporating electricity price signals and carbon emission factors is constructed. This generates energy storage charging and discharging scheduling strategies and orderly charging instructions for electric vehicles on a minute-by-minute timescale. This process is subject to node voltage constraints and transformer capacity limitations in the distribution network. Deep reinforcement learning algorithms are used to extract features from historical operating data, gradually improving the system's adaptability to uncertainties on both the source and load sides. Ultimately, this system achieves dynamic energy supply and demand balance, optimal economic performance, and low-carbon operation.

[0003] When dealing with the strong randomness and multi-time-scale coupling characteristics of the source and load sides, the high-dimensional nonlinear constraints of the mixed integer programming model make it difficult for the solver to converge to the feasible domain within a finite time window. In particular, when the fluctuations in renewable energy output and the sudden changes in electric vehicle charging demand form spatiotemporal correlation interference, the traditional rolling optimization framework is prone to fall into the local optimal solution, resulting in a mismatch between the control instruction sequence and the real-time operating conditions, thereby weakening the dynamic regulation robustness of the system. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an energy management and control system and method for integrating electric vehicles and renewable energy, which is used to solve the problem of rolling optimization path deviation caused by the difficulty in solving high-dimensional mixed integer programming models and the interference of spatiotemporal correlation in existing control and regulation systems.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] The first aspect is the energy management and control system for electric vehicles and renewable energy integration, including:

[0007] An acquisition module is used to collect photovoltaic output data, wind power output data, and electric vehicle load demand data through a distributed sensor network, filter outliers in the data using an anomaly detection algorithm, perform timing compensation on multi-source data based on a timing alignment algorithm, and output synchronized timing data to the optimization control module;

[0008] The optimization control module is used to receive the synchronous time series data output by the acquisition module and generate optimization instructions in layers according to long-term, medium-term and short-term time scales:

[0009] The long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand based on the synchronized time series data, dynamically modifies the energy storage capacity threshold in combination with the energy storage life loss model, and inputs the modified threshold into the medium-term optimization layer;

[0010] The medium-term optimization layer embeds a robust optimization model in the model predictive control framework, generates rolling optimization instructions based on the modified energy storage capacity threshold and the real-time electricity price signal, and transmits the instructions to the short-term optimization layer;

[0011] The short-term optimization layer uses a distributed solver to decompose the mixed integer programming problem in the rolling optimization instruction and generates charging and discharging instructions based on the voltage constraint priority of the distribution network node;

[0012] A control execution module is used to receive the charge and discharge instructions generated by the short-term optimization layer, issue the instructions to the electric vehicle bidirectional converter through a multi-priority message queue, dynamically adjust the execution threshold based on stability control theory, and feed back the execution results to the constraint module;

[0013] A constraint module, configured to update the policy network parameters in the optimization control module through a dual experience replay mechanism, decouple the spatiotemporal correlation constraints of the charge and discharge instructions in combination with causal reasoning technology, and input the decoupled constraints into the distributed solver of the short-term optimization layer;

[0014] A fault-tolerant module is used to simulate extreme operating condition data through an adversarial sample generation network, inject the policy network training data set of the optimization control module, analyze the root cause of the instruction conflict of the control execution module based on the Bayesian network, and dynamically adjust the service priority of the mid-term optimization layer.

[0015] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand through a Monte Carlo sampling method, and dynamically corrects the charge and discharge capacity thresholds of the energy storage unit based on the energy storage life loss model;

[0016] The mid-term optimization layer receives the scenario tree generated by the long-term optimization layer, dynamically allocates multi-objective optimization weights using the entropy weight method, generates rolling optimization instructions including electricity price signals and carbon emission factors, and transmits the instructions to the short-term optimization layer;

[0017] The short-term optimization layer decomposes the mixed integer programming problem into parallel subtasks using a topological sorting algorithm based on the rolling optimization instructions transmitted by the medium-term optimization layer and the voltage constraint priority of the distribution network nodes, and assigns the subtasks to the distributed solver to perform parallel computing.

[0018] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the constraint module includes:

[0019] A federated learning unit is used to encrypt the electric vehicle load data on the edge side using differential privacy technology, upload the encrypted local model parameters to the blockchain node for distributed storage, and generate global model parameters through federated aggregation, which are then sent to the policy network of the optimization control module;

[0020] The constraint decomposition unit receives the global model parameters generated by the federated learning unit, smoothes the feasible domain boundary of the mixed integer programming problem based on the hyperplane projection algorithm, generates dynamic constraint conditions and inputs them into the distributed solver of the short-term optimization layer.

[0021] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the fault-tolerant module includes:

[0022] a fault-tolerant training unit, configured to train the policy network of the optimization control module in stages through a curriculum learning strategy, and inject extreme operating condition data generated by a generative adversarial network into the training data set of the dual experience replay mechanism of the optimization control module to enhance the anti-interference capability of the policy network;

[0023] The service orchestration unit dynamically adjusts the optimization layer service priority according to the resource availability evaluation results. When the resource availability is lower than the preset threshold, the scene tree generation service of the long-term optimization layer is blocked, and a downgrade instruction is sent to the medium-term optimization layer to switch to a simplified rolling optimization model.

[0024] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the acquisition module includes:

[0025] A spectrum analysis unit is used to perform spectrum analysis on photovoltaic output data, identify the main frequency characteristics of periodic noise components, and dynamically match the basis function type and decomposition level of the wavelet denoising algorithm according to the main frequency characteristics to generate denoised photovoltaic output data;

[0026] The timing alignment unit receives the denoised photovoltaic output data and the original wind power output data generated by the spectrum analysis unit, calibrates the high-frequency collected electric vehicle load data and the low-frequency updated renewable energy forecast data through a cache queue with timestamp matching, generates synchronized timing data, and outputs it to the long-term optimization layer and the medium-term optimization layer of the optimization control module.

[0027] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the closed-loop verification unit of the control execution module models the dependency relationship of charging and discharging instructions through a directed acyclic graph. When an instruction conflict is detected, the characteristic parameters of the conflicting instructions are transmitted to the Bayesian network reasoning unit of the fault-tolerant module to locate the root cause of the conflict and generate adjustment instructions.

[0028] The instruction issuing unit retransmits the invalid instruction according to the verification failure result, and feeds back the verification log including the conflict type and the number of retransmissions to the policy distillation unit of the constraint module for updating the training priority of the policy network in the optimization control module.

[0029] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the medium-term optimization layer receives the energy storage capacity threshold generated by the long-term optimization layer through a rolling time domain window, generates charge and discharge power instructions on a minute-level time scale in combination with real-time electricity price signals, and transmits the instructions to the short-term optimization layer;

[0030] The short-term optimization layer calculates the electric vehicle cluster scheduling plan that meets the distribution network node voltage constraints based on the received charging and discharging power instructions through the distributed solver, marks the instructions that exceed the voltage safety threshold, and inputs the marked out-of-limit instructions into the constraint decomposition unit of the constraint module to adjust the dynamic constraint conditions.

[0031] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the strategy learning unit of the constraint module migrates the constraint parsing rules of the complex strategy network in the optimization control module to a lightweight model through knowledge migration technology, and deploys the migrated model parameters to the edge-side controller;

[0032] The federated learning unit of the constraint module detects the parameter conflict of the lightweight model based on the parameter version information stored in the blockchain, triggers a parameter rollback operation to restore the historical valid parameters, and synchronizes the rollback instruction to all edge nodes to update the local model parameters.

[0033] Furthermore, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the consistency unit of the fault-tolerant module smoothes the parameter update sequence of the policy network in the optimization control module using a parameter smoothing algorithm, and inputs the processed parameters into the causal reasoning unit of the constraint module to decouple the spatiotemporal correlation constraints in the charge and discharge instructions;

[0034] The fault-tolerant training unit of the fault-tolerant module dynamically adjusts the difficulty gradient of the course learning strategy based on the simulation data of historical extreme working conditions, and feeds back the gradient adjustment signal to the generative adversarial network to optimize the generation parameters of the adversarial samples.

[0035] In a second aspect, the present invention provides an energy management and control method for integrating electric vehicles and renewable energy, which is applied to the energy management and control system for integrating electric vehicles and renewable energy, and includes:

[0036] The acquisition module collects photovoltaic output data, wind power output data, and electric vehicle load demand data, uses the isolation forest algorithm to filter outliers, and uses the dynamic time warping algorithm to compensate for time series deviations to generate synchronized time series data.

[0037] By optimizing the long-term optimization layer of the control module, an uncertainty scenario tree of renewable energy output and load demand is generated based on the synchronized time series data, and the energy storage capacity threshold is dynamically corrected in combination with the energy storage life loss model;

[0038] The mid-term optimization layer of the optimization control module receives the energy storage capacity threshold value corrected by the long-term optimization layer, embeds the robust optimization layer in the model predictive control framework, and generates rolling optimization instructions in combination with the real-time electricity price signal;

[0039] By using the short-term optimization layer of the optimization control module, a distributed solver is used to decompose the mixed integer programming problem in the rolling optimization instruction, and the voltage constraints of the distribution network nodes are processed first to generate charging and discharging instructions;

[0040] The charging and discharging instructions are issued to the bidirectional converter device of the electric vehicle through the control execution module, the execution threshold is dynamically adjusted based on the Lyapunov stability theory, and the execution result is fed back to the constraint module;

[0041] The policy network parameters of the optimization control module are updated through the dual experience replay mechanism of the constraint module, the spatiotemporal correlation constraints in the charge and discharge instructions are decoupled using causal reasoning technology, and the decoupled constraints are input into the distributed solver of the short-term optimization layer;

[0042] The generative adversarial network of the fault-tolerant module simulates extreme working conditions to generate adversarial samples and injects them into the training data set of the policy network. Based on the Bayesian network inference of the root cause of the instruction conflict, the service priority of the mid-term optimization layer is dynamically adjusted.

[0043] Beneficial effects of the present invention:

[0044] The beneficial effect of the present invention is that it significantly reduces the solution complexity of high-dimensional mixed integer programming models through multi-time scale hierarchical optimization architecture and dynamic constraint decomposition technology. The long-term optimization layer compresses the solution space dimension based on the uncertainty scenario tree. The medium-term optimization layer uses robust optimization and dynamic weight distribution to generate anti-interference instructions. The short-term optimization layer uses a distributed solver to parallelly process voltage constraint problems, thereby improving solution efficiency. It combines causal reasoning and federated learning mechanisms to decouple spatiotemporal correlation constraints, suppress path deviations caused by new energy output fluctuations and load mutations, and the closed-loop verification module dynamically corrects instruction conflicts through directed acyclic graphs and Bayesian networks. The fault-tolerant module adopts curriculum learning strategies and adversarial sample training to enhance the system's anti-interference ability. The acquisition module's timing alignment and noise suppression technology ensure the consistency of multi-source data, ultimately achieving dynamic balance of energy supply and demand, optimized instruction robust execution, and efficient cross-layer resource collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solution of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on the drawings without paying any creative labor.

[0046] Figure 1 This is a flow chart of an energy management and control method for integrating electric vehicles and renewable energy provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings.

[0048] In order to better understand the purpose of the present invention, the present invention is described in further detail below.

[0049] First, the energy management and control system for electric vehicles and renewable energy integration provided by this company includes:

[0050] An acquisition module is configured to collect photovoltaic output data, wind power output data, and electric vehicle load demand data through a distributed sensor network, filter outliers in the data using an isolation forest algorithm, perform time series alignment on multi-source data based on a dynamic time warping algorithm, and output the processed multi-source data to an optimization control module;

[0051] The optimization control module is used to receive the multi-source data output by the acquisition module and generate optimization instructions based on time scales:

[0052] The long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand based on the multi-source data, modifies the energy storage capacity threshold based on the energy storage life loss model, and outputs it to the medium-term optimization layer;

[0053] The medium-term optimization layer generates rolling optimization instructions by embedding the model predictive control framework of the robust optimization layer, combining the energy storage capacity threshold and the real-time electricity price signal, and outputs them to the short-term optimization layer;

[0054] The short-term optimization layer uses a distributed alternating direction multiplier method solver to decompose the mixed integer programming problem in the rolling optimization instructions, giving priority to processing the voltage constraints of the distribution network nodes to generate charging and discharging instructions;

[0055] a control execution module, configured to receive the charge and discharge instructions generated by the short-term optimization layer, issue the instructions to the electric vehicle bidirectional converter through a multi-priority message queue, dynamically adjust the instruction execution threshold based on Lyapunov stability theory, and feed back the execution results to the constraint module;

[0056] A constraint module is used to update the policy network parameters in the optimization control module through a dual experience replay mechanism, decouple the spatiotemporal correlation constraints in the charge and discharge instructions using causal reasoning technology, and input the constraint decomposition results into the distributed solver of the short-term optimization layer;

[0057] A fault-tolerant module is used to simulate extreme working conditions and inject them into the training data set of the policy network through a generative adversarial network, and to dynamically downgrade the mid-term optimization layer service priority of the optimization control module based on the Bayesian network reasoning of the root cause of the instruction conflict in the control execution module.

[0058] The energy management and control system for integrating electric vehicles and renewable energy, provided by the present invention, collects photovoltaic output data, wind power output data, and electric vehicle load demand data in real time through a distributed sensor network. It uses an isolation forest algorithm to detect and filter outliers in the collected data, and uses a dynamic time warping algorithm to compensate and align the timing deviations of multi-source data. This synchronized time series data is then transmitted to an optimization control module. The optimization control module processes data at different time scales. The long-term optimization layer uses Monte Carlo sampling to generate an uncertainty scenario tree for renewable energy output and load demand. It dynamically corrects the capacity threshold of the energy storage unit using an energy storage life loss model and outputs this to the medium-term optimization layer. The medium-term optimization layer embeds a robust optimization layer within a model predictive control framework, generating rolling optimization instructions based on real-time electricity price signals and the corrected energy storage capacity threshold, and transmits these instructions to the short-term optimization layer. The short-term optimization layer uses a distributed alternating direction multiplier method solver to decompose the mixed integer programming problem, prioritizing the voltage constraints of distribution network nodes to generate charging and discharging instructions.

[0059] After receiving the charge and discharge instructions from the short-term optimization layer, the control execution module distributes them hierarchically to the electric vehicle's bidirectional converter via a multi-priority message queue. Simultaneously, it dynamically adjusts instruction execution thresholds based on Lyapunov stability theory to ensure stability during the charging and discharging process. Feedback data generated during execution is transmitted to the constraint module, which updates the parameters of the policy network in the optimization control module using a dual experience replay mechanism. This module utilizes causal reasoning techniques to decouple the spatiotemporal constraints in the charge and discharge instructions. These decomposed dynamic constraints are then fed into the distributed solver in the short-term optimization layer to optimize subsequent instruction generation. The fault-tolerance module simulates extreme operating conditions using a generative adversarial network and injects the policy network's training dataset to enhance model robustness. Furthermore, the module uses a Bayesian network to reason about the root causes of instruction conflicts within the control execution module and dynamically adjusts the service priorities of the medium-term optimization layer to address unexpected disruptions.

[0060] The acquisition module further includes an edge-side fast Fourier transform unit, which is used to identify periodic noise components in photovoltaic output data and dynamically adjust the number of decomposition layers of the wavelet denoising algorithm based on the noise spectrum characteristics to improve data preprocessing accuracy. The timing alignment unit calibrates the high-frequency collected electric vehicle load data and the low-frequency updated renewable energy forecast data through a cache queue to achieve synchronization of multi-source data. The medium-term optimization layer of the optimization control module combines the energy storage capacity threshold output by the long-term optimization layer with the real-time electricity price signal within a rolling time domain window to generate charging and discharging power instructions on a minute time scale. Based on this instruction, the short-term optimization layer calculates the electric vehicle cluster scheduling plan that meets the node voltage constraints, marks the out-of-limit instructions, and inputs them into the constraint decomposition unit for dynamic constraint adjustment.

[0061] The constraint module uses policy distillation to transfer knowledge from complex policy networks to a lightweight model, which is then deployed to edge controllers to reduce computational load. The federated learning unit utilizes differential privacy to protect edge data, maintains traceability of model parameters through blockchain-based authentication, and triggers a rollback mechanism to synchronize all edge nodes when parameter conflicts are detected. The consistency unit of the fault-tolerance module uses an exponentially weighted moving average algorithm to smooth parameter fluctuations in the policy network, ensuring stability in the optimization process. Furthermore, the difficulty gradient of the course learning strategy is adjusted based on historical extreme operating condition simulation results, and the adversarial example generation logic is optimized to enhance the system's anti-interference capabilities.

[0062] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand through Monte Carlo sampling, and dynamically adjusts the charge and discharge capacity thresholds of the energy storage unit based on the energy storage life loss model;

[0063] The mid-term optimization layer inputs the scenario tree into a multi-objective weight allocation model dynamically calculated by the entropy weight method, generates a rolling optimization instruction including an electricity price signal and a carbon emission factor, and transmits the instruction to the short-term optimization layer;

[0064] The short-term optimization layer prioritizes the distribution network node voltage constraints through topological sorting according to the rolling optimization instructions, and distributes the decomposed sub-problems to the distributed alternating direction multiplier method solver for parallel calculation.

[0065] In the energy management and control system for integrating electric vehicles and renewable energy described in the present invention, the long-term optimization layer simulates the uncertainty distribution of renewable energy output and load demand through the Monte Carlo sampling method, and generates an uncertainty scenario tree containing multi-time section power fluctuation characteristics. Based on the mapping relationship between the number of charge and discharge cycles and the capacity attenuation rate in the energy storage life loss model, the long-term optimization layer dynamically adjusts the charge and discharge capacity threshold of the energy storage unit to avoid the impact of overcharging and over-discharging on the life of the energy storage equipment, and outputs the corrected capacity threshold to the medium-term optimization layer. The medium-term optimization layer receives the capacity threshold and the real-time electricity price signal, calculates the dynamic weights of the economic goal and the low-carbon goal in the rolling optimization through the entropy weight method, constructs a multi-objective optimization function including the electricity price sensitivity coefficient and the carbon emission factor, and combines with the robust optimization layer to generate rolling optimization instructions to combat prediction errors, which are transmitted to the short-term optimization layer as the constraint boundary.

[0066] After receiving the rolling optimization instructions from the mid-term optimization layer, the short-term optimization layer identifies nodes at risk of voltage overshooting based on the distribution network topology. Using a topological sorting algorithm, it prioritizes charging and discharging power allocation within voltage-sensitive regions. The short-term optimization layer decomposes the mixed integer programming problem into multiple subproblems and computes them in parallel using a distributed alternating direction multiplier method solver. The optimization results of each subproblem are synchronized through a consistency coordination mechanism to generate charging and discharging instructions for the electric vehicle cluster that satisfy the node voltage constraints. The distributed solver marks the overshoot instructions as dynamic constraint adjustment signals, which are fed back to the constraint decomposition unit to update the feasible region boundaries in the short-term optimization model. Parameter transfer and constraint inheritance form a closed-loop optimization chain between the optimization layers, enabling the gradual refinement and dynamic coordination of instructions across multiple timescales.

[0067] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the constraint module includes: a federated learning unit for protecting electric vehicle load data at the edge using differential privacy technology, uploading encrypted local model parameters to a blockchain node for evidence storage, and sending federated aggregated global model parameters to the policy network in the optimization control module;

[0068] The constraint decomposition unit smoothes the feasible domain boundary through a hyperplane projection algorithm based on the global model parameters updated by the federated learning unit, generates dynamic constraint conditions and inputs them into the distributed alternating direction multiplier method solver of the short-term optimization layer.

[0069] In the energy management and control system for integrating electric vehicles and renewable energy, a federated learning unit uses differential privacy technology to add noise and perturb data collected by the edge EV load data, generating a local dataset that meets privacy protection requirements. This local dataset is fed into the edge controller to train the policy network submodel, generating encrypted model parameters with timestamps and device identification information. These parameters are then uploaded to a distributed ledger node for storage via a blockchain smart contract. The federated learning unit aggregates the encrypted model parameters from multiple edge nodes, using a homomorphic encryption algorithm to fuse the parameters and update the weights. This global model parameter set is then distributed to the optimization control module's policy network, enabling collaborative model optimization across edge nodes.

[0070] After receiving the global model parameters updated by the federated learning unit, the constraint decomposition unit analyzes the spatiotemporal correlation constraints output by the policy network and smoothes the feasible region boundaries of the mixed integer programming problem using a hyperplane projection algorithm. Based on the current distribution network topology and real-time load distribution, the constraint decomposition unit identifies the coupling relationship between voltage and power balance constraints, decomposes the complex constraints into independent sub-constraint sets, and generates dynamic adjustment coefficients, which are input into the distributed alternating direction multiplier method solver in the short-term optimization layer. The distributed solver adjusts the iteration step size and penalty factor based on the dynamic constraints, synchronously updating the constraint boundaries of each sub-problem during parallel computing, reducing the dimensionality of the optimization model and improving solution efficiency.

[0071] The federated learning unit and the constraint decomposition unit form a closed-loop feedback mechanism through parameter transfer. Updates to global model parameters trigger the constraint decomposition unit to reassess the boundaries of the feasible region. Adjustments to dynamic constraints further influence the training data distribution of the policy network. Model version information stored on the blockchain provides historical baseline parameters for constraint decomposition. When parameter conflicts are detected, the federated learning unit invokes a smart contract to verify model version consistency and trigger a parameter rollback to maintain the stability of the optimization process. Upon receiving the rollback instruction, the edge node extracts historically valid parameters from the blockchain distributed ledger to replace the anomalous parameters, ensuring the reliability of the constraint decomposition unit's input data.

[0072] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the fault-tolerant module includes:

[0073] A fault-tolerant training unit, configured to gradually improve the anti-interference capability of the policy network in the optimization control module through a curriculum learning strategy, and to inject extreme operating condition data simulated by the generative adversarial network into the training data set of the dual experience replay mechanism;

[0074] The service orchestration unit dynamically blocks the services of the long-term optimization layer according to the resource availability evaluation result, and transmits the degradation instruction to the mid-term optimization layer to adjust the generation logic of the rolling optimization instruction.

[0075] In the energy management and control system for integrated electric vehicles and renewable energy, the fault-tolerant training unit constructs a training scenario library in stages based on a curriculum learning strategy. Initially, the policy network is trained using historical data from normal operating conditions. Extreme power fluctuations and load mutation scenarios simulated by a generative adversarial network are gradually introduced to enhance the policy network's adaptability to abnormal operating conditions. The generative adversarial network's generator reconstructs data from extreme events, such as grid voltage drops and sudden drops in renewable energy output, using a spatiotemporal feature extraction module. The discriminator dynamically adjusts the credibility threshold of the generated data, combining the distribution characteristics of historical real-world data with a dual experience replay mechanism. This data is then injected into the policy network's training dataset to enhance model robustness.

[0076] The service orchestration unit monitors computing resource utilization and communication link status in real time. It uses the Analytic Hierarchy Process (AHP) to assess the computational load of the long-term optimization layer and the real-time requirements of the mid-term optimization layer. When resource availability falls below a preset threshold, it dynamically disables the scenario tree generation service in the long-term optimization layer and simultaneously sends a downgrade instruction to the mid-term optimization layer. Upon receiving the downgrade instruction, the mid-term optimization layer switches to a simplified model predictive control framework, employing a fixed-weight multi-objective optimization model instead of a dynamic weight allocation mechanism. This shortens the rolling optimization instruction generation cycle and maintains a minute-by-minute instruction issuance frequency.

[0077] The fault-tolerant training unit and the service orchestration unit operate collaboratively via a data bus. Generative adversarial network simulations of extreme operating conditions trigger changes in resource availability assessment indicators. The service orchestration unit dynamically adjusts the phase switching speed of the curriculum learning strategy based on the complexity of the operating conditions. As the policy network's anti-interference capabilities improve, the service orchestration unit gradually restores some service functions in the long-term optimization layer. Using a progressive weight migration method, simplified model parameters are integrated into the full optimization framework, achieving a smooth transition across service levels. Blockchain nodes record operation logs during service degradation and recovery, providing time-stamped state transition data for policy network training and optimizing the data sampling strategy for the dual experience replay mechanism.

[0078] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the acquisition module includes:

[0079] an edge-side fast Fourier transform unit, configured to identify periodic noise components of photovoltaic power output data and dynamically adjust the number of decomposition layers of a wavelet denoising algorithm according to noise spectrum characteristics to preprocess the photovoltaic power output data;

[0080] The timing alignment unit calibrates the high-frequency collected electric vehicle load data and the low-frequency updated renewable energy forecast data through a cache queue, and outputs the synchronized timing data to the long-term optimization layer and the medium-term optimization layer of the optimization control module.

[0081] In the energy management and control system for integrated electric vehicles and renewable energy, an edge-side fast Fourier transform (FFT) unit performs spectral analysis on photovoltaic output data, extracting periodic noise characteristics through fundamental frequency detection and harmonic component identification. Based on the amplitude and distribution characteristics of the dominant frequency components in the noise spectrum, the wavelet denoising algorithm dynamically matches the basis function type and adaptively adjusts the number of decomposition levels to isolate noise frequency bands while preserving the low-frequency trend components of the photovoltaic output signal. The denoised photovoltaic output data is then merged with the original wind power output data and fed into a time series alignment unit for multi-source data fusion.

[0082] The time series alignment unit establishes a timestamp-based cache queue mechanism for the frequently collected electric vehicle load data and the less frequently updated renewable energy forecast data. High-frequency data is sliced ​​into fixed time windows and temporarily stored in the queue. The arrival of low-frequency data triggers matching and interpolation calculations for the corresponding time segments within the queue. Missing time point data is filled in using a linear weighted compensation algorithm to generate a synchronized time series dataset. Synchronized data is resampled according to the timescale requirements of the long-term and medium-term optimization layers. The long-term optimization layer receives hourly trend data for scenario tree construction, while the medium-term optimization layer obtains minute-level detailed data to support rolling optimization.

[0083] The edge-side fast Fourier transform (FT) unit and the timing alignment unit form a serial processing chain. The denoised photovoltaic data output by the FT unit carries noise marker information. The timing alignment unit uses this marker information to weight and de-weight abnormal periods in the high-frequency load data, suppressing the interference of noise propagation on the optimization control module. The optimization control module's feedback signal dynamically adjusts the length of the spectral analysis window of the FT unit. When increasing fluctuations in renewable energy output are detected, the analysis window is shortened to improve the timeliness of noise identification. The interpolation calculation frequency of the timing alignment unit is simultaneously optimized to maintain the time alignment accuracy of multi-source data.

[0084] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the closed-loop verification unit of the control execution module describes the dependency relationship of charging and discharging instructions through a directed acyclic graph, and when an instruction conflict is detected, the Bayesian network reasoning unit of the fault-tolerant module is triggered to locate the source of the conflict.

[0085] The instruction issuing unit retransmits the failed instruction according to the cyclic redundancy check result, and feeds back the check log to the policy distillation unit of the constraint module to optimize the policy network parameters.

[0086] In the energy management and control system for integrating electric vehicles and renewable energy described in the present invention, a closed-loop verification unit establishes a dependency model of charging and discharging instructions based on a directed acyclic graph, maps each charging and discharging instruction into a graph node, and maps the timing constraints and resource competition relationships between instructions into directed edges. The closed-loop verification unit monitors the execution status of the nodes in the graph in real time. When resource preemption or time window overlap between nodes is detected, it is marked as an instruction conflict event, triggering the Bayesian network reasoning unit of the fault-tolerant module. The Bayesian network receives the priority label, current resource occupancy rate, and historical conflict data of the conflicting instructions, and determines through probabilistic reasoning that the root cause of the conflict is an abnormal resource allocation strategy or too tight time scale coupling, and generates an instruction priority adjustment plan or a resource release instruction.

[0087] The instruction dispatch unit embeds a cyclic redundancy check (CRC) code when transmitting charge and discharge instructions. If the receiving end fails the verification, it locates the physical link node of the failed instruction based on the instruction timestamp and device identification information. Based on the hierarchical division of the multi-priority message queue, it prioritizes the retransmission of high-priority voltage over-limit correction instructions. A verification log records the content, failure type, and number of retransmissions of failed instructions. It then appends a timestamp and environmental parameters and transmits it to the policy distillation unit of the constraint module. This unit analyzes the conflict pattern characteristics in the log, extracts the spatiotemporal distribution of frequently failed instructions, adjusts the sampling weights of historical data in the dual experience replay mechanism, and optimizes the training data distribution of the lightweight policy network.

[0088] The closed-loop verification unit and the policy distillation unit interact with each other to form a dynamic tuning mechanism. The conflict root cause analysis results output by the Bayesian network are encoded as constraint adjustment signals and fed into the policy distillation unit's parameter update process, guiding the lightweight model to enhance its generalization capabilities for specific conflict scenarios. After the optimized policy network parameters are deployed to the edge controller, new charge and discharge instruction dependency data is generated, and the structural features of the directed acyclic graph are reversely updated to reduce the triggering frequency of similar conflict events. The fault-tolerance module synchronously receives the policy network's update log and dynamically adjusts the extreme operating condition simulation parameters of the generative adversarial network to ensure that adversarial examples cover the policy network's optimized instruction generation pattern.

[0089] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the mid-term optimization layer of the optimization control module receives the energy storage capacity threshold generated by the long-term optimization layer through a rolling time domain window, and generates charge and discharge power instructions on a minute-level time scale in combination with real-time electricity price signals;

[0090] The short-term optimization layer calculates the electric vehicle cluster scheduling plan that meets the node voltage constraint based on the charge and discharge power instructions through the distributed alternating direction multiplier method solver, and marks the over-limit instructions and inputs them into the constraint decomposition unit of the constraint module.

[0091] In the energy management and control system for integrated electric vehicles and renewable energy described in this invention, the medium-term optimization layer receives the energy storage capacity threshold output by the long-term optimization layer within a rolling time-domain window and constructs a dynamic optimization objective function based on real-time electricity price signals. The rolling time-domain window dynamically adjusts its window length based on the statistical characteristics of the renewable energy output forecast error. A robust optimization layer is embedded within the model predictive control framework to generate charging and discharging power instructions that are robust against forecast fluctuations. By introducing slack variables and uncertainty bound constraints, the robust optimization layer maps real-time electricity price signals into economic weight coefficients for power regulation, generating minute-by-minute rolling optimization instructions that are output to the short-term optimization layer.

[0092] After receiving these rolling optimization instructions, the short-term optimization layer identifies voltage over-limit risks in high-load areas based on the distribution network topology and node voltage sensitivity analysis. A distributed alternating direction multiplier method solver decomposes the mixed integer programming problem into multiple sub-optimization tasks. Through parallel computation, it coordinates the power allocation solutions for each sub-problem to generate charging and discharging scheduling instructions for the electric vehicle cluster that meet voltage constraints. The solver marks instructions that exceed voltage safety thresholds, attaches timestamps and node location information, and then inputs them into the constraint decomposition unit of the constraint module.

[0093] The constraint decomposition unit analyzes spatiotemporal constraints based on marked out-of-bounds instructions. It adjusts the relaxation factors of the feasible region boundaries using a hyperplane projection algorithm to generate dynamic constraint adjustment coefficients. These adjusted constraints are fed back to the distributed solver in the short-term optimization layer to update the constraint boundaries and penalty parameters of the sub-optimization tasks, forming a closed-loop optimization mechanism. The constraint module simultaneously receives the spatiotemporal distribution characteristics of the out-of-bounds instructions and uses a dual experience replay mechanism to update the training data priority of the policy network, optimizing the generation logic of subsequent rolling optimization instructions.

[0094] The mid-term and short-term optimization layers form a cross-timescale collaborative mechanism through instruction flow and constraint feedback. The optimization results of the rolling time window constrain the feasible domain of the short-term optimization. Out-of-limit instructions in the short-term optimization layer trigger dynamic corrections to the scenario tree parameters in the long-term optimization layer. The parallel computation results of the distributed solver reversely update the robust optimization parameters within the rolling time window, improving the mid-term optimization layer's adaptability to real-time operating condition fluctuations and maintaining the dynamic balance of the multi-level optimization process.

[0095] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the policy learning unit of the constraint module transfers the knowledge of the complex policy network to a lightweight model through the policy distillation technology, and deploys the distilled model parameters to the edge-side controller;

[0096] The federated learning unit rolls back the update operation that conflicts with the lightweight model based on the parameter version information stored in the blockchain, and synchronizes the rollback instruction to all edge nodes to update the local model.

[0097] In the energy management and control system for integrated electric vehicles and renewable energy described in the present invention, the policy learning unit constructs a teacher-student model architecture through policy distillation technology, migrating the temporal and spatial correlation constraint knowledge implicit in the complex policy network to a lightweight model. The teacher-student model uses an attention mechanism to match feature layer outputs. By jointly optimizing the soft label loss function and the feature distillation loss function, it extracts the key parameters of the charge and discharge instruction generation pattern in the complex policy network, generating a lightweight model suitable for edge computing resources. The distilled model parameters are quantized, compressed, and formatted before being deployed to the real-time inference engine of the edge controller for rapid generation of local charge and discharge instructions.

[0098] The federated learning unit uses a blockchain smart contract to record the model parameter version information of each edge node, including parameter hash values, update timestamps, and device identifiers. When a gradient conflict is detected between the newly uploaded local parameters and the global parameters of the lightweight model, the federated learning unit invokes the smart contract to verify the consistency of the parameter versions and trigger a rollback operation based on the version tree. The rollback instruction carries the blockchain index of the historically valid parameters and is synchronized to all edge nodes via a peer-to-peer communication protocol. The conflicting parameters are replaced and the lightweight model is reloaded, maintaining global consistency during the distributed optimization process.

[0099] The policy learning unit and the federated learning unit form a collaborative optimization mechanism via a data bus. The deployment parameters of the lightweight model are fed back into the training process of the complex policy network, enhancing the original model's adaptability to edge computing scenarios through adversarial training. Parameter update logs stored on the blockchain are parsed as time series features and fed into the policy distillation loss function to dynamically adjust knowledge transfer weights and optimize the feature alignment accuracy of the teacher and student models. Upon receiving the rollback instruction, the edge node synchronously updates the constraint parsing rules of the local model. Through incremental learning, historically valid parameters are integrated with real-time data to reduce the impact of parameter conflicts on the optimization control module.

[0100] Specifically, in the energy management and control system for integrating electric vehicles and renewable energy according to the present invention, the consistency unit of the fault-tolerant module smoothes parameter mutations in the strategy network of the optimization control module using an exponentially weighted moving average algorithm, and inputs the smoothed parameters into the causal reasoning unit of the constraint module to decouple spatiotemporal constraints.

[0101] The fault-tolerant training unit adjusts the difficulty gradient of the course learning strategy based on the simulation results of historical extreme working conditions, and feeds the adjustment signal back to the generative adversarial network to update the adversarial sample generation logic.

[0102] In the energy management and control system for integrating electric vehicles and renewable energy described in the present invention, the consistency unit of the fault-tolerant module smoothes the parameter update sequence of the strategy network in the optimization control module through an exponentially weighted moving average algorithm, thereby suppressing the violent fluctuations of the parameters caused by fluctuations in the output of new energy or sudden changes in load. The smoothed parameters carry the stability characteristics of the historical training phase and are input to the causal reasoning unit of the constraint module to analyze the coupling relationship between the voltage regulation target and the charge and discharge power distribution in the spatiotemporal constraints, and generate a set of independent constraints after decoupling. The causal reasoning unit adjusts the projection direction of the constraint decomposition unit based on the decoupling result, reduces the dimensional conflict of the feasible domain boundary, and improves the convergence efficiency of the distributed solver of the short-term optimization layer.

[0103] The fault-tolerant training unit reconstructs a compound fault scenario involving a sudden drop in renewable energy output and a surge in electric vehicle load based on historical extreme operating condition data stored on the blockchain. Using a curriculum learning strategy, the training difficulty gradient is gradually increased. Initially, a single fault mode data set is loaded to train the policy network. Temporal and spatial correlation interference factors are gradually added. The generative adversarial network dynamically adjusts the intensity and complexity of the extreme condition simulation based on the policy network's current interference tolerance. The signal feedback to the generative adversarial network's spatiotemporal feature extraction module is adjusted, optimizing the noise injection ratio of the generator and the anomaly detection threshold of the discriminator to generate adversarial examples covering multiple dimensions of interference features.

[0104] The consistency unit and the fault-tolerant training unit form a dynamic equilibrium mechanism through parameter interaction. The smoothing parameters output by the exponentially weighted moving average algorithm serve as feature inputs to the generative adversarial network, constraining the generation range of adversarial examples to match the adaptation boundary of the policy network. Adjustments to the difficulty gradient of the curriculum learning strategy trigger the consistency unit to recalculate the parameter smoothing coefficients, optimizing the decoupling accuracy of the causal inference unit. The fault-tolerant module simultaneously records the generated features of adversarial examples and the response data of the policy network. The federated learning unit updates the blockchain-stored fault mode library, providing incremental data support for subsequent parameter smoothing and adversarial training.

[0105] Second, see Figure 1 The energy management and control method for integrating electric vehicles and renewable energy provided by the present invention is applied to the energy management and control system for integrating electric vehicles and renewable energy as described above, including:

[0106] Step S101: Collect photovoltaic output data, wind power output data, and electric vehicle load demand data, use the isolation forest algorithm to filter outliers, and compensate for time series deviations based on the dynamic time warping algorithm, and output the processed multi-source data;

[0107] Step S102: Generate optimization instructions based on the processed multi-source data at different time scales:

[0108] Long-term optimization generates uncertainty scenario trees for renewable energy output and load demand, and corrects energy storage capacity thresholds based on energy storage life loss models;

[0109] The mid-term optimization generates rolling optimization instructions by combining the energy storage capacity threshold and real-time electricity price signals through a model predictive control framework embedded in a robust optimization layer;

[0110] Short-term optimization utilizes a distributed alternating direction multiplier method solver to decompose the mixed integer programming problem in the rolling optimization instructions, giving priority to processing the voltage constraints of the distribution network nodes to generate charging and discharging instructions;

[0111] Step S103, issuing the charge and discharge instructions to the electric vehicle bidirectional converter device through a multi-priority message queue, dynamically adjusting the instruction execution threshold based on Lyapunov stability theory, and feeding back the execution results to the strategy network parameter update process;

[0112] Step S104, updating the policy network parameters in the optimization control module through a dual experience replay mechanism, decoupling the spatiotemporal correlation constraints in the charge and discharge instructions using causal reasoning technology, and inputting the constraint decomposition results into the distributed solver of the short-term optimization;

[0113] Step S105, generating adversarial samples by simulating extreme working conditions through a generative adversarial network and injecting them into the training data set of the policy network, and dynamically downgrading the service priority of the mid-term optimization layer based on the Bayesian network reasoning instruction conflict root cause.

[0114] The technical features of the present invention are explained as follows:

[0115] Acquisition module:

[0116] Distributed sensor network: It consists of multiple sensors distributed in photovoltaic power stations, wind farms and electric vehicle charging stations, and is used to collect real-time photovoltaic output, wind power output and electric vehicle load data.

[0117] Isolation Forest Algorithm: An unsupervised machine learning algorithm used to detect and filter outliers in data to improve data quality.

[0118] Dynamic Time Warping (DTW) is a time series alignment technique used to compensate for time deviations between multi-source data with different acquisition frequencies (such as high-frequency load data and low-frequency forecast data) and generate synchronized time series data.

[0119] Optimized control module:

[0120] Long-term optimization layer: Generates an uncertainty scenario tree of renewable energy output and load demand based on Monte Carlo sampling, simulates power fluctuations in multiple time sections in the future, and dynamically corrects the energy storage capacity threshold in combination with the energy storage life loss model to avoid overcharging and over-discharging.

[0121] Mid-term optimization layer: A robust optimization model is embedded in the model predictive control (MPC) framework, and rolling optimization instructions are generated in combination with real-time electricity price signals. The entropy weight method is used to dynamically allocate the weights of economic efficiency and low-carbon targets to improve the anti-interference ability of instructions.

[0122] Short-term optimization layer: Using the distributed alternating direction method of multipliers (ADMM) solver, the mixed integer programming problem is decomposed into parallel subtasks, the voltage constraints of the distribution network nodes are prioritized, and the charging and discharging instructions are generated.

[0123] Control execution module:

[0124] Multi-priority message queue: Instructions are sent to the electric vehicle bidirectional converter according to their urgency (e.g., voltage over-limit correction instructions have the highest priority), enabling priority execution of key instructions.

[0125] Lyapunov stability theory: Maintain grid operation stability and avoid voltage or frequency oscillations by dynamically adjusting instruction execution thresholds (such as charge and discharge power limits).

[0126] Constraint Module:

[0127] Dual experience replay mechanism: Sample and train the policy network from historical data and real-time feedback, balance the weights of new and old experience, and avoid model overfitting.

[0128] Causal reasoning technology: The hyperplane projection algorithm is used to decouple the spatiotemporal correlation constraints in the charging and discharging instructions (such as the coupling relationship between voltage regulation and power distribution) to generate the dynamic feasible region boundary.

[0129] Fault-tolerant module:

[0130] Generative Adversarial Network (GAN): Consists of a generator and a discriminator. The generator simulates extreme operating conditions (such as a sudden drop in renewable energy output), while the discriminator verifies the authenticity of the data and enhances the anti-interference ability of the policy network.

[0131] Bayesian Network: Uses probabilistic reasoning to locate the root causes of instruction conflicts (such as resource contention or timing conflicts) and generate priority adjustment plans or resource release instructions.

[0132] Monte Carlo sampling: Generate multiple sets of possible renewable energy output and load demand scenarios through random sampling, build an uncertainty scenario tree, and compress the solution space for long-term optimization.

[0133] Federated learning: While protecting edge data privacy (through differential privacy encryption), it aggregates local model parameters of multiple edge nodes to generate a global optimization model.

[0134] Blockchain evidence storage: Encrypted model parameters and operation logs are stored in a distributed ledger to ensure that data cannot be tampered with and provide a traceability basis for parameter version conflicts.

[0135] Policy distillation technology: Migrates the knowledge of complex policy networks to lightweight models, reducing the computational load of edge controllers while retaining the ability to parse key constraints.

[0136] Curriculum learning strategy: Train the policy network in stages, gradually increasing the complexity of adversarial samples (such as from single faults to compound faults) to enhance the model's generalization ability.

[0137] The present invention collects synchronous time series data of the module → long-term optimization layer scene tree → mid-term optimization layer rolling instructions → short-term optimization layer charge and discharge instructions → control execution module issuance and feedback → constraint module dynamic adjustment → fault-tolerant module conflict diagnosis to form a closed-loop optimization link.

[0138] Cross-layer collaboration: The long-term optimization layer provides mid-term optimization based on energy storage capacity threshold constraints, the mid-term optimization layer generates short-term optimization based on instruction boundary constraints, and the short-term optimization layer provides feedback on over-limit instructions to correct long-term scenario tree parameters, achieving dynamic coordination at multiple time scales.

[0139] Anti-interference mechanism: The fault-tolerant module improves the robustness of the policy network through adversarial sample training, the Bayesian network diagnoses conflicts in real time, and the service orchestration unit dynamically downgrades non-critical services to ensure the stability of the system under extreme working conditions.

[0140] Long-term optimization layer model:

[0141] Monte Carlo sampling model:

[0142] Function: Generates uncertainty scenario trees for renewable energy (photovoltaic, wind power) output and electric vehicle load demand through random sampling, simulating power fluctuation distribution in multiple time sections.

[0143] Technical role: Provides a multi-dimensional solution space for long-term optimization and reduces the complexity of high-dimensional mixed integer programming problems.

[0144] Mid-term optimization layer model:

[0145] Model Predictive Control (MPC) model embedded with robust optimization:

[0146] Function: Combines real-time electricity price signals and energy storage capacity thresholds within a rolling time window to generate rolling optimization instructions that are resistant to forecast errors.

[0147] Technical role: By introducing slack variables and uncertainty boundary constraints, the robustness of the command to fluctuations in renewable energy output is improved.

[0148] Short-term optimization layer model:

[0149] Distributed Alternating Direction Method of Multipliers (ADMM) solver

[0150] Function: Decompose the mixed integer programming problem into parallel subtasks, prioritize the voltage constraints of distribution network nodes, and generate charging and discharging instructions.

[0151] Technical function: Identify voltage-exceeding risk nodes through topological sorting, coordinate the optimization results of each sub-problem, and improve solution efficiency.

[0152] Constraint module model:

[0153] Federated Learning Model:

[0154] Function: Aggregate local model parameters of edge nodes to generate a global model while protecting data privacy (differential privacy encryption).

[0155] Technical function: Parameter version traceability is achieved through blockchain evidence storage, parameter rollback is triggered after conflict detection, and optimization consistency is maintained.

[0156] Causal Inference Models:

[0157] Function: Analyze the spatiotemporal coupling relationship in charge and discharge instructions (such as the correlation between voltage regulation and power distribution) and generate a decoupled independent constraint set.

[0158] Technical function: Suppress the deviation of the optimization path caused by fluctuations in new energy and sudden changes in load, and improve the stability of instruction execution.

[0159] Fault-tolerant module model:

[0160] Generative Adversarial Network (GAN) Model:

[0161] Function: The generator simulates extreme operating conditions (such as grid voltage drop and sudden drop in renewable energy output), and the discriminator verifies the authenticity of the data.

[0162] Technical function: Injecting adversarial samples into the policy network training dataset to enhance the model's adaptability to abnormal working conditions.

[0163] Bayesian network model:

[0164] Function: Locates the root causes of instruction conflicts (such as resource contention and timing overlap) based on probabilistic reasoning and generates priority adjustment or resource release instructions.

[0165] Technical function: Cooperate with the closed-loop verification module to dynamically correct instruction dependencies and reduce the frequency of conflict triggering.

[0166] Acquisition module model:

[0167] Dynamic Time Warping (DTW) algorithm:

[0168] Function: Calibrate high-frequency collected electric vehicle load data and low-frequency updated renewable energy forecast data to generate synchronized time series data.

[0169] Technical function: Through timestamp matching and interpolation compensation, it eliminates the timing deviation of multi-source data and provides consistent input for hierarchical optimization.

[0170] Wavelet denoising model:

[0171] Function: Identify the periodic noise components of photovoltaic output data based on spectrum analysis, dynamically adjust the wavelet basis function and decomposition level, and retain the effective signal.

[0172] Technical function: Suppress the interference of noise on optimization control and improve the accuracy of data preprocessing.

[0173] Policy Network Model:

[0174] Double experience playback mechanism:

[0175] Function: Balance sampling from historical data and real-time feedback, update policy network parameters, and avoid model overfitting.

[0176] Technical role: Combined with policy distillation technology, complex policy knowledge is migrated to a lightweight model to reduce the computing load on the edge side.

[0177] Service orchestration model:

[0178] Analytical Hierarchy Process (AHP) Resource Assessment Model:

[0179] Function: Dynamically evaluate computing resource utilization and communication link status, and generate service priority adjustment signals.

[0180] Technical function: When resource availability falls below the threshold, non-critical services in the long-term optimization layer are blocked, and mid-term optimization is switched to a simplified model to ensure real-time performance.

[0181] Specific embodiments of the present invention relate to an energy management and control system and method for integrating electric vehicles and renewable energy. This system addresses the command mismatch and path deviation issues encountered in existing technologies when solving high-dimensional mixed-integer programming and addressing spatiotemporal correlation interference. This system utilizes a hierarchical optimization architecture and dynamic constraint decomposition technology to achieve efficient control. The system includes an acquisition module, an optimization control module, a control execution module, a constraint module, and a fault-tolerance module. Specific implementations are as follows:

[0182] The acquisition module collects real-time output data from photovoltaic power plants, wind farms, and electric vehicle charging stations through a distributed sensor network. It uses the isolation forest algorithm to detect and filter outliers in the raw data, removing anomalous data points. For periodic noise in the photovoltaic output data, a fast Fourier transform is used to perform spectral analysis. After identifying the dominant frequency characteristics of the noise, the wavelet denoising algorithm dynamically matches the basis function type and the number of decomposition layers, setting the number of decomposition layers to 3-5, effectively separating the noise frequency bands while retaining low-frequency trend components. For high-frequency electric vehicle load data and low-frequency renewable energy forecast data, a dynamic time warping algorithm is used to align the time series. A timestamp-matched cache queue is used to calibrate data timing deviations, generating synchronized time series data that is output to the long-term and medium-term optimization layers of the optimization control module.

[0183] After receiving the synchronized time series data, the long-term optimization layer of the optimization control module uses the Monte Carlo sampling method to generate an uncertainty scenario tree for renewable energy output and load demand, simulating hourly fluctuations in renewable energy output and changes in load demand over the next 24 hours, and generating 100-200 sets of scenario samples. Combined with the energy storage life loss model, the charge and discharge capacity thresholds of the energy storage unit are dynamically corrected based on the mapping relationship between the number of charge and discharge cycles and the capacity attenuation rate. The correction range is ±5% of the nominal capacity to avoid the impact of overcharging and over-discharging on the life of the energy storage equipment, and the corrected capacity threshold is passed to the medium-term optimization layer. The medium-term optimization layer embeds a robust optimization model in the model predictive control framework, with a rolling time domain window length set to 15-30 minutes. Combining real-time electricity price signals and the corrected energy storage capacity threshold, the entropy weight method is used to dynamically allocate weights for economic objectives and low-carbon objectives, generating rolling optimization instructions containing electricity price sensitivity coefficients and carbon emission factors, and transmitting them to the short-term optimization layer. The short-term optimization layer uses a distributed alternating direction multiplier method solver to decompose the mixed integer programming problem into 8-12 parallel subtasks. It identifies voltage-limit risk nodes based on the distribution network topology, prioritizes the charging and discharging power allocation problem in sensitive areas through topological sorting, generates electric vehicle cluster scheduling instructions that meet node voltage constraints, marks instructions that exceed the voltage safety threshold of ±5%, and feeds back the limit-limit instructions to the constraint module.

[0184] After receiving the charge and discharge commands generated by the short-term optimization layer, the control execution module distributes them to the electric vehicle's bidirectional converter via a multi-priority message queue. Priority is determined by the urgency of the voltage constraints, with voltage over-limit correction commands being assigned the highest priority. The command execution threshold is dynamically adjusted based on Lyapunov stability theory, with the charge and discharge power adjustment step set between 2% and 5% of the rated power to ensure a stable charging and discharging process. Execution results, including command execution success rate, voltage fluctuation range, and power deviation data, are transmitted to the constraint module via a closed-loop feedback link. The constraint module uses a dual experience replay mechanism to update the policy network parameters, sampling from a historical data pool and real-time feedback data in a 7:3 ratio. Causal reasoning techniques are used to analyze the spatiotemporal coupling relationships within the charge and discharge commands. Complex constraints are decomposed into independent sub-constraint sets using a hyperplane projection algorithm. The relaxation factor at the feasible region boundary is dynamically adjusted to a range of 0.1-0.3. Constraint conditions are generated and fed into the distributed solver in the short-term optimization layer to optimize the subsequent command generation logic.

[0185] The fault-tolerance module simulates extreme operating conditions such as grid voltage drops, sudden drops in renewable energy output, and load surges using a generative adversarial network. The generator, taking historical fault data as input, reconstructs adversarial examples with voltage drops of 10%-20% of the rated value and output drops of 5%-10% / minute. The discriminator verifies the sample's credibility based on the distribution characteristics of real data in a dual experience replay mechanism. Adversarial examples that meet threshold criteria are injected into the policy network's training dataset at a ratio of 20%-30%. A Bayesian network receives instruction conflict characteristic parameters transmitted by the closed-loop verification module, including priority labels, resource utilization, and historical conflict frequencies. Using probabilistic reasoning, it locates the root cause of the conflict as abnormal resource allocation or overlapping time windows. It then generates instruction priority adjustment plans or resource release instructions, dynamically reducing the service priority of the mid-term optimization layer to 70%-80% of its original priority. This triggers the service orchestration unit to disable the scenario tree generation service of the long-term optimization layer, switching the mid-term optimization layer to a simplified model predictive control framework. This adopts a fixed-weight multi-objective optimization model instead of dynamic weight allocation, shortening the rolling optimization instruction generation cycle to 50% of its original cycle.

[0186] During system operation, the federated learning unit uses differential privacy technology to add Laplace noise (noise scale set to 0.1-0.5) to edge-side electric vehicle load data. The encrypted local model parameters are timestamped and device-identified before being uploaded to blockchain nodes for distributed storage. Edge node parameters are aggregated every 1-2 hours, fused using a homomorphic encryption algorithm to generate global model parameters, and distributed to the policy network. If a conflict is detected between a parameter hash value and a historical version, the smart contract is triggered to roll back to the most recent valid version, synchronously updating the local model parameters of all edge nodes with a tolerance of ±2%. The policy distillation unit migrates the constraint parsing rules in the complex policy network to a lightweight model, compressing the model parameters to 30%-40% of the original model. This model is then deployed to the edge controller for real-time inference, with response latency controlled to less than 200 milliseconds.

[0187] Through the above implementation methods, the system realizes the dynamic generation and closed-loop correction of multi-time-scale instructions. The long-term optimization layer compresses the solution space dimension, the medium-term optimization layer improves the instruction anti-interference ability, and the short-term optimization layer efficiently solves the voltage constraint problem. The constraint decomposition and fault-tolerant mechanism are combined to suppress path deviation, and ultimately achieve the goals of energy supply and demand balance, optimal economy and low-carbon operation.

[0188] This invention effectively addresses the difficulty of solving high-dimensional mixed integer programming models through a multi-time-scale hierarchical optimization architecture and dynamic constraint decomposition technology. The long-term optimization layer of the optimization control module generates an uncertainty scenario tree based on Monte Carlo sampling, and combines it with the energy storage life loss model to compress the solution space dimensions. The medium-term optimization layer embeds a robust optimization layer within the model predictive control framework, dynamically assigning multi-objective weights through the entropy weight method to generate rolling instructions, reducing the complexity of the real-time optimization problem. The short-term optimization layer uses a distributed alternating direction multiplier method solver to decompose the mixed integer programming into parallel subproblems, and uses topological sorting to prioritize voltage limit constraints, achieving efficient solution of high-dimensional models.

[0189] To address path deviations caused by spatiotemporal correlation interference, the constraint module decouples spatiotemporal constraints using causal reasoning techniques and integrates a federated learning mechanism to update policy network parameters. The causal reasoning unit analyzes the spatiotemporal coupling relationships in charge and discharge instructions, smoothing the boundaries of the feasible region using a hyperplane projection algorithm to generate dynamic constraints. The federated learning unit aggregates encrypted parameters from edge nodes to construct a global model. Using blockchain evidence storage, it tracks parameter version conflicts and rolls back abnormal updates to eliminate the impact of spatiotemporal interference on the optimized path.

[0190] The system maintains the stability of the rolling optimization path through a closed-loop feedback mechanism. The control execution module verifies instruction dependencies based on a directed acyclic graph, triggering a Bayesian network to locate the source of conflicts and retransmit failed instructions. The fault-tolerant training unit of the fault-tolerant module employs a curriculum learning strategy to inject adversarial examples in stages, dynamically adjusting the extreme operating condition simulation logic of the generative adversarial network. The acquisition module uses a dynamic time warping algorithm to compensate for timing deviations in multi-source data, providing precise input for hierarchical optimization. These modules collaborate to achieve closed-loop generation and dynamic correction of multi-objective optimization instructions, mitigating the risk of path deviation.

Claims

1. An energy management and control system integrating electric vehicles and renewable energy, characterized by: include: The acquisition module is used to collect photovoltaic output data, wind power output data and electric vehicle load demand data, process the collected data, and output synchronous time series data to the optimization control module; The optimization control module is used to receive synchronous time series data and generate optimization instructions according to the long-term, medium-term, and short-term time scales: The long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand based on the synchronized time series data, dynamically modifies the energy storage capacity threshold in combination with the energy storage life loss model, and inputs the modified threshold into the medium-term optimization layer; The medium-term optimization layer embeds a robust optimization model in the model predictive control framework, generates rolling optimization instructions based on the modified energy storage capacity threshold and the real-time electricity price signal, and transmits the rolling optimization instructions to the short-term optimization layer; The short-term optimization layer uses a distributed solver to decompose the mixed integer programming problem in the rolling optimization instruction and generates charging and discharging instructions based on the voltage constraint priority of the distribution network node; A control execution module is used to receive the charge and discharge instructions generated by the short-term optimization layer, issue the instructions to the electric vehicle bidirectional converter through a multi-priority message queue, dynamically adjust the execution threshold based on stability control theory, and feed back the execution results to the constraint module; A constraint module, configured to update the policy network parameters in the optimization control module through a dual experience replay mechanism, decouple the spatiotemporal correlation constraints of the charge and discharge instructions in combination with causal reasoning technology, and input the decoupled constraints into the distributed solver of the short-term optimization layer; A fault-tolerant module is used to simulate extreme operating condition data through an adversarial sample generation network, inject the policy network training data set of the optimization control module, analyze the root cause of the instruction conflict of the control execution module based on the Bayesian network, and dynamically adjust the service priority of the mid-term optimization layer.

2. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The long-term optimization layer generates an uncertainty scenario tree of renewable energy output and load demand through the Monte Carlo sampling method, and dynamically corrects the charge and discharge capacity thresholds of the energy storage unit based on the energy storage life loss model; The mid-term optimization layer receives the scenario tree generated by the long-term optimization layer, dynamically allocates multi-objective optimization weights using the entropy weight method, generates rolling optimization instructions including electricity price signals and carbon emission factors, and transmits the instructions to the short-term optimization layer; The short-term optimization layer decomposes the mixed integer programming problem into parallel subtasks using a topological sorting algorithm based on the rolling optimization instructions transmitted by the medium-term optimization layer and the voltage constraint priority of the distribution network nodes, and assigns the subtasks to the distributed solver to perform parallel computing.

3. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The constraint module includes: A federated learning unit is used to encrypt the electric vehicle load data on the edge side using differential privacy technology, upload the encrypted local model parameters to the blockchain node for distributed storage, and generate global model parameters through federated aggregation, which are then sent to the policy network of the optimization control module; The constraint decomposition unit receives the global model parameters generated by the federated learning unit, smoothes the feasible domain boundary of the mixed integer programming problem based on the hyperplane projection algorithm, generates dynamic constraint conditions and inputs them into the distributed solver of the short-term optimization layer.

4. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The fault-tolerant module includes: a fault-tolerant training unit, configured to train the policy network of the optimization control module in stages through a curriculum learning strategy, and inject extreme operating condition data generated by a generative adversarial network into the training data set of the dual experience replay mechanism of the optimization control module to enhance the anti-interference capability of the policy network; The service orchestration unit dynamically adjusts the optimization layer service priority according to the resource availability evaluation results. When the resource availability is lower than the preset threshold, the scene tree generation service of the long-term optimization layer is blocked, and a downgrade instruction is sent to the medium-term optimization layer to switch to a simplified rolling optimization model.

5. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The acquisition module includes: A spectrum analysis unit is used to perform spectrum analysis on photovoltaic output data, identify the main frequency characteristics of periodic noise components, and dynamically match the basis function type and decomposition level of the wavelet denoising algorithm according to the main frequency characteristics to generate denoised photovoltaic output data; The timing alignment unit receives the denoised photovoltaic output data and the original wind power output data generated by the spectrum analysis unit, calibrates the high-frequency collected electric vehicle load data and the low-frequency updated renewable energy forecast data through a cache queue with timestamp matching, generates synchronized timing data, and outputs it to the long-term optimization layer and the medium-term optimization layer of the optimization control module.

6. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The closed-loop verification unit of the control execution module models the dependency relationship of the charge and discharge instructions through a directed acyclic graph. When an instruction conflict is detected, the characteristic parameters of the conflicting instructions are transmitted to the Bayesian network reasoning unit of the fault-tolerant module to locate the root cause of the conflict and generate an adjustment instruction. The instruction issuing unit retransmits the invalid instruction according to the verification failure result, and feeds back the verification log including the conflict type and the number of retransmissions to the strategy distillation unit of the constraint module for updating the training priority of the strategy network in the optimization control module.

7. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The mid-term optimization layer receives the energy storage capacity threshold generated by the long-term optimization layer through a rolling time domain window, generates a minute-level charge and discharge power instruction based on the real-time electricity price signal, and transmits the instruction to the short-term optimization layer; The short-term optimization layer calculates the electric vehicle cluster scheduling plan that meets the distribution network node voltage constraints based on the received charging and discharging power instructions through the distributed solver, marks the instructions that exceed the voltage safety threshold, and inputs the marked out-of-limit instructions into the constraint decomposition unit of the constraint module to adjust the dynamic constraint conditions.

8. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The strategy learning unit of the constraint module migrates the constraint parsing rules of the complex strategy network in the optimization control module to a lightweight model through knowledge migration technology, and deploys the migrated model parameters to the edge-side controller; The federated learning unit of the constraint module detects the parameter conflict of the lightweight model based on the parameter version information stored in the blockchain, triggers a parameter rollback operation to restore the historical valid parameters, and synchronizes the rollback instruction to all edge nodes to update the local model parameters.

9. The energy management and control system for integrating electric vehicles and renewable energy according to claim 1, characterized in that: The consistency unit of the fault-tolerant module smoothes the parameter update sequence of the policy network in the optimization control module through a parameter smoothing algorithm, and inputs the processed parameters into the causal reasoning unit of the constraint module for decoupling the spatiotemporal correlation constraints in the charge and discharge instructions; The fault-tolerant training unit of the fault-tolerant module dynamically adjusts the difficulty gradient of the course learning strategy based on the simulation data of historical extreme working conditions, and feeds back the gradient adjustment signal to the generative adversarial network to optimize the generation parameters of the adversarial samples.

10. An energy management and control method for integrating electric vehicles and renewable energy, applied to an energy management and control system for integrating electric vehicles and renewable energy as claimed in any one of claims 1 to 9, characterized in that: include: The acquisition module collects photovoltaic output data, wind power output data, and electric vehicle load demand data, uses the isolation forest algorithm to filter outliers, and uses the dynamic time warping algorithm to compensate for time series deviations to generate synchronized time series data. By optimizing the long-term optimization layer of the control module, an uncertainty scenario tree of renewable energy output and load demand is generated based on the synchronized time series data, and the energy storage capacity threshold is dynamically corrected in combination with the energy storage life loss model; The mid-term optimization layer of the optimization control module receives the energy storage capacity threshold value corrected by the long-term optimization layer, embeds the robust optimization layer in the model predictive control framework, and generates rolling optimization instructions in combination with the real-time electricity price signal; By using the short-term optimization layer of the optimization control module, a distributed solver is used to decompose the mixed integer programming problem in the rolling optimization instruction, and the voltage constraints of the distribution network nodes are processed first to generate charging and discharging instructions; The charging and discharging instructions are issued to the bidirectional converter device of the electric vehicle through the control execution module, the execution threshold is dynamically adjusted based on the Lyapunov stability theory, and the execution result is fed back to the constraint module; The policy network parameters of the optimization control module are updated through the dual experience replay mechanism of the constraint module, the spatiotemporal correlation constraints in the charge and discharge instructions are decoupled using causal reasoning technology, and the decoupled constraints are input into the distributed solver of the short-term optimization layer; The generative adversarial network of the fault-tolerant module simulates extreme working conditions to generate adversarial samples and injects them into the training data set of the policy network. Based on the Bayesian network inference of the root cause of the instruction conflict, the service priority of the mid-term optimization layer is dynamically adjusted.

Images