A wind-solar-storage multi-site coordinated intelligent optimization scheduling method

Abstract

This invention discloses a collaborative intelligent optimization scheduling method for multiple wind, solar, energy storage, and charging power plants, relating to the fields of power system automation and new energy dispatching. The method constructs a two-level cloud-edge collaborative architecture and a three-layer time-scale scheduling framework of "day-ahead, intraday, and real-time," establishes a nonlinear decay model of the storage system's state of equilibrium (SOH) coupled with multiple factors, builds a five-dimensional multi-objective dynamic weight optimization system, and adopts a hierarchical distributed solution: the day-ahead layer achieves fair collaboration among multiple stakeholders through asymmetric Nash negotiation and ADMM; the intraday layer uses distributed MPC rolling correction; and the real-time layer employs a MADRL-MPC hybrid algorithm with an attention mechanism to balance decision-making efficiency and constraint security. This invention can improve the renewable energy absorption rate, extend the lifespan of energy storage, and ensure data privacy and fair benefits for multiple stakeholders. It is applicable to the comprehensive energy optimization operation of virtual power plants and industrial park microgrids.

Description

Technical Field

[0001] This invention relates to the fields of power system automation, comprehensive utilization of new energy and distributed intelligent dispatching, and in particular to a collaborative intelligent optimization dispatching method for multiple wind, solar, energy storage and charging stations based on cloud-edge collaboration, multi-time-scale coupling and multi-entity game theory. It is applicable to the comprehensive energy optimization operation of virtual power plants, park-level microgrid groups and new energy aggregation areas that include wind farms, photovoltaic power generation stations, electrochemical energy storage stations and electric vehicle charging stations. Background Technology

[0002] Driven by both the "dual carbon" goals and the construction of a new power system, the installed capacity of new energy power generation such as wind power and photovoltaics continues to climb. Electrochemical energy storage and electric vehicle charging facilities are rapidly becoming widespread as flexible regulation resources, gradually forming a regional energy system with multiple power stations clustered together in the "wind-solar-storage-charging" model. Through coordinated and optimized scheduling of multiple power stations, fluctuations in new energy output can be mitigated, the absorption rate of new energy can be improved, the peak-valley difference in the power grid can be reduced, and the economic efficiency of system operation can be enhanced. This is the core technological support for building a new power system.

[0003] Currently, some progress has been made in the technical research and engineering application of coordinated wind, solar, energy storage, and charging scheduling. However, there are still systemic deficiencies in multi-timescale coordination, energy storage lifetime management, intelligent algorithm implementation, coordination of interests among multiple stakeholders, handling of uncertainties, and low-carbon integration. These deficiencies are specifically reflected in the following aspects:

[0004] First, the lack of coordination mechanisms in multi-timescale dispatching easily leads to command conflicts and scheme failures. Power system dispatching encompasses three timescales: day-ahead planning, intraday rolling correction, and real-time control. The dispatching objectives and response speeds at different levels differ fundamentally. Existing technologies mostly employ single-layer dispatching models, or, while layered, lack constraint transmission and deviation correction mechanisms between layers. This results in frequent adjustments at the real-time layer causing the energy storage SOC to rapidly deviate from the day-ahead planning benchmark, rendering the day-ahead economic dispatching scheme completely infeasible. This leads to problems such as control command oscillations and logic deadlocks, and in severe cases, even threatens the stable operation of the power grid. Some layered schemes focus only on centralized dispatching by a single entity, failing to adapt to scenarios with widely distributed multi-site facilities, resulting in high communication pressure and insufficient deployment flexibility.

[0005] Second, the modeling of energy storage lifespan degradation is crude and not deeply coupled with upper-level scheduling. Most existing scheduling models simplify battery energy storage as an ideal energy conversion unit, only constraining the upper and lower limits of charge and discharge power and the SOC operating range, severely neglecting the electrochemical mechanisms of battery aging. The cycle life of lithium-ion batteries is strongly coupled with multiple factors such as depth of discharge, charge / discharge rate, operating temperature, and SOC dwell time, and includes both cycle aging and calendar aging degradation mechanisms. Existing technologies mostly use simplified throughput methods or single DOD empirical formulas, which cannot accurately quantify capacity degradation under different operating conditions. Scheduling algorithms cannot achieve an optimal balance between short-term gains and long-term lifespan, resulting in the actual lifespan of energy storage systems being far lower than the design value, significantly increasing the total lifespan investment cost.

[0006] Third, intelligent scheduling algorithms struggle to balance constraint safety and interpretability, limiting their engineering implementation. While deep reinforcement learning-based intelligent scheduling methods possess excellent dynamic adaptability and real-time response capabilities, their "black box" nature prevents operators from understanding the decision-making rationale and easily leads to the output of illegal commands that violate physical hard constraints such as equipment capacity and grid power flow, failing to meet the core requirements of "observable, interpretable, and controllable" power system scheduling. Traditional model predictive control, while strictly satisfying constraints, lacks sufficient global optimization capabilities and struggles to handle high-dimensional nonlinear multi-station collaborative scenarios. Existing technologies have failed to achieve a deep integration of these two approaches, thus failing to simultaneously meet the dual requirements of global economy and operational safety.

[0007] Fourth, multi-stakeholder collaboration faces privacy barriers and imbalances in benefit distribution. Wind, solar, energy storage, and charging stations are often operated by different entities, resulting in low data interoperability. Centralized scheduling schemes require each station to upload all core operational data, posing a risk of data privacy breaches. Furthermore, the profit-driven nature of each entity leads to insufficient willingness to collaborate. Existing distributed scheduling schemes only achieve power coordination and lack a fair and reasonable benefit distribution mechanism. Entities making significant contributions to regulation do not receive corresponding benefits, easily leading to "passive resistance" and failing to fully leverage the scale regulation advantages of multi-station aggregation. Some game-theoretic allocation methods adopt an average distribution principle, failing to consider the differences in capacity and regulation flexibility among entities, which does not conform to actual business logic.

[0008] Fifth, the system has weak uncertainty handling capabilities, making it difficult to balance economic efficiency and security. Wind and solar power output is highly random and volatile due to weather conditions, and electric vehicle charging demand also exhibits significant spatiotemporal uncertainty. Existing forecasting methods have large errors, directly causing scheduling decisions to deviate from the optimal state, leading to wind and solar power curtailment or grid power exceeding limits. Robust optimization schemes, in pursuit of safety in extreme scenarios, set overly conservative uncertainty sets, resulting in significantly increased system operating costs and low resource utilization; while stochastic optimization schemes suffer from insufficient scenario representativeness and high computational complexity, making it difficult to meet the real-time requirements of engineering projects.

[0009] Sixth, the integration of low-carbon goals is insufficient, and policy adaptability needs to be improved. Existing dispatch methods mostly focus on minimizing operating costs as the core objective, and rarely organically integrate policy factors such as carbon trading costs and clean energy consumption responsibility weights into the collaborative dispatch model. This weakens the guiding role of green dispatch and fails to help power plants participate in the carbon market and green certificate trading to obtain additional revenue, resulting in insufficient alignment with the "dual carbon" goals.

[0010] Therefore, there is an urgent need to develop a collaborative intelligent optimization scheduling method for wind, solar, energy storage and charging multi-sites that can achieve seamless connection across multiple time scales, precise control of energy storage lifespan, secure and explainable intelligent algorithms, fair and equitable collaboration among multiple stakeholders, efficient handling of uncertainties, and integration of low-carbon goals, in order to solve the aforementioned systemic problems existing in the current technology. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a collaborative intelligent optimization scheduling method for multiple wind, solar, energy storage and charging stations. This method solves the technical problems in the prior art, such as scheduling conflicts at multiple time scales, rough energy storage life modeling, imbalance between the constraints and interpretability of intelligent algorithms, difficulty in balancing the privacy and interests of multiple stakeholders, conservative and inefficient handling of uncertainties, and insufficient integration of low-carbon goals. This method achieves global optimization of multi-station resources, safe and economical operation, and mutual benefit and win-win results for multiple stakeholders.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: a collaborative intelligent optimization scheduling method for multiple wind, solar, energy storage, and charging stations, comprising the following steps:

[0013] S1. Cloud-edge collaborative architecture construction and multi-source heterogeneous data preprocessing and multi-scale prediction: Construct a two-level cloud-edge collaborative architecture of "cloud dispatch center - site edge gateway" to perform topology identification and parameter collection for all wind farms, photovoltaic power generation stations, electrochemical energy storage stations and electric vehicle charging stations in the region, and clean, normalize and spatiotemporally align multi-source heterogeneous operation data; construct the STGCN-Transformer hybrid prediction model to generate probability prediction intervals for wind and solar power output and charging load at three time scales: day-ahead, intraday, and real-time, providing data support for hierarchical scheduling.

[0014] S2. Construction of a three-layer multi-timescale collaborative scheduling framework: A three-layer scheduling architecture is built, consisting of "day-ahead global optimization layer - intraday rolling correction layer - real-time rapid response layer". The time resolution, optimization objectives, decision variables and constraint boundaries of each layer are clearly defined, and a benchmark transfer and deviation feedback mechanism between upper and lower layers is established. The day-ahead layer is deployed in the cloud, while the intraday layer and the real-time layer are deployed on the field edge gateway to achieve seamless connection and distributed deployment of scheduling at different time scales.

[0015] S3. Multi-factor coupled nonlinear degradation modeling of energy storage SOH: Distinguishing between the dual degradation mechanisms of cyclic aging and calendar aging, the rainflow counting method is used to decompose the charge-discharge cycle. Combining the Arrhenius temperature effect equation and the charge-discharge rate penalty function, a multi-factor coupled energy storage battery health state degradation model is constructed to quantify the capacity degradation caused by a single charge-discharge process and convert it into lifetime loss cost and incorporate it into the optimization objective.

[0016] S4. Construction of a multi-objective dynamic weight optimization model: A five-dimensional multi-objective optimization function is established with the core objectives of minimizing the overall system operating cost, maximizing the renewable energy absorption rate, minimizing energy storage life loss, minimizing power fluctuation at the grid connection point, and minimizing carbon trading costs. A dynamic weight adjustment mechanism integrating subjective and objective factors is constructed based on the entropy weight method and fuzzy hierarchical analysis method, and the weight coefficients of each optimization objective are automatically adapted according to the real-time operating scenario.

[0017] S5. Layered Distributed Intelligent Optimization Solution: Differentiated solution algorithms are adopted to address the different characteristics of the three-layer scheduling architecture: The day-ahead layer is based on an asymmetric Nash negotiation game to construct a distributed collaborative optimization model and uses the alternating direction multiplier method to solve the problem, achieving fair distribution of benefits among multiple stakeholders while protecting data privacy; the intraday layer uses distributed model predictive control for rolling correction to track day-ahead plan deviations; the real-time layer uses the MADRL-MPC hybrid algorithm with an attention mechanism to achieve fast decision-making, hard constraint fallback, and interpretable output.

[0018] S6. Dispatch instruction issuance and closed-loop feedback of operation status: Dispatch instructions generated at each level are decomposed and issued to the corresponding stations for execution. Real-time operation data of each station is collected, and the deviation between the actual operation status and the planned value is calculated. The deviation is fed back to the intraday rolling correction layer and the real-time response layer. An online incremental learning mechanism is established to continuously optimize the prediction model and scheduling algorithm, so as to realize the closed-loop iterative optimization of the scheduling process.

[0019] Furthermore, the STGCN-Transformer hybrid prediction model described in step S1 is constructed by abstracting the multi-site system into a graph structure. Where V is the set of station nodes, E is the set of edges, and A is the adjacency matrix reflecting the electrical distance and geographical correlation between stations; firstly, spatial topological features and short-term temporal dependencies are extracted through a spatiotemporal graph convolutional network, and the calculation formula for the gated temporal convolutional layer is:

[0020]

[0021] In the formula, The input features are for the l-th layer. This represents the graph convolution operation. W is a learnable parameter. The Sigmoid activation function is used; the extracted spatial feature sequence is input into the Transformer encoder, and the long-distance time series dependence is captured through the self-attention mechanism; finally, a deterministic prediction benchmark of 24 hours (1-hour resolution) before the day, a 95% confidence probability interval of 4 hours (15-minute resolution) within the day, and a real-time 15-minute (1-minute resolution) ultra-short-term rolling prediction value are generated.

[0022] Furthermore, the specific configuration of the three-layer scheduling architecture described in step S2 is as follows:

[0023] The current global optimization layer is deployed in the cloud-based dispatch center and is executed at 24:00 every day. The optimization cycle is 24 hours and the time resolution is 1 hour. With the goal of optimizing the overall efficiency of the system throughout the entire cycle, it outputs the 24-hour baseline power plan and energy storage SOC baseline range for each site as a hard constraint for the lower-level dispatch.

[0024] Intraday Rolling Correction Layer: Deployed at the site edge gateway, it is executed every 15 minutes, with an optimization cycle of 4 hours and a time resolution of 15 minutes. The goal is to minimize the deviation of the day-ahead plan. It combines the latest forecast data to correct power commands, correct SOC deviation, and reserve 10% of the rated power for adjustment.

[0025] Real-time rapid response layer: Deployed at the edge gateway of the power station, it executes once every 1 minute, with an optimization cycle of 5 minutes and a time resolution of 1 minute. With the goal of smoothing real-time power fluctuations and meeting power grid safety constraints, it outputs millisecond-level adjustment commands based on a hybrid intelligent algorithm.

[0026] The upper and lower layer constraint transmission mechanism is as follows: the power plan and SOC range output by the upper layer serve as hard constraints for the lower layer, and the optimization of the lower layer must not exceed the boundary set by the upper layer; when the cumulative SOC deviation exceeds 5% of the rated capacity, the rolling correction of the upper layer plan is triggered, forming a collaborative mechanism of "baseline transmission - deviation correction - closed-loop feedback".

[0027] Furthermore, the specific construction process of the nonlinear decay model of the energy storage SOH mentioned in step S3 is as follows:

[0028] 1. Cyclic Aging Calculation: The SOC fluctuation sequence within the scheduling period is cyclically decomposed using the rainflow counting method to obtain N independent charge-discharge cycles. The discharge depth of the i-th cycle is then extracted. Average operating temperature With charge / discharge rate The capacity decay per cycle is:

[0029]

[0030] In the formula, These are the fitting coefficients for the battery material properties. The activation energy is given by R, where R is the ideal gas constant. This is a multiplier penalty function, which grows exponentially when the multiplier exceeds the rated value;

[0031] 2. Calendar Aging Calculation: Calendar aging is related to runtime, ambient temperature, and average SOC. The calculation formula is as follows:

[0032]

[0033] In the formula, Calendar aging factor. For calendar aging activation energy, The average temperature over the period. Let SOC stay interval influence function;

[0034] 3. Total Attenuation and Cost Calculation: The total SOH attenuation during the scheduling period is... Converted into lifespan depreciation cost:

[0035]

[0036] In the formula, The initial investment cost for energy storage, This is the end-of-life threshold (usually 0.8).

[0037] Furthermore, the specific form of the multi-objective optimization function in step S4 is as follows:

[0038]

[0039] The sub-objectives are defined as follows:

[0040] Overall system operating cost This includes grid purchase and sale costs, energy storage operation and maintenance costs, wind and solar curtailment penalty costs, and carbon trading costs. The calculation formula is as follows:

[0041]

[0042] Among them, carbon trading costs , Carbon emission intensity of power grid supply, Free carbon emission allowances;

[0043] New energy curtailment rate Minimize the proportion of wind and solar power curtailment in total power generation and maximize the level of power consumption;

[0044] Energy storage lifespan loss rate Minimize the ratio of cumulative energy storage capacity decay to rated capacity within the scheduling cycle;

[0045] Power fluctuation at grid connection point Minimize the fluctuation amplitude and rate of change of active power at the grid connection point to meet the grid connection assessment requirements;

[0046] 5. Loss of user charging satisfaction Minimize the proportion of unmet charging needs to ensure a good user charging experience.

[0047] The dynamic weight adjustment mechanism is as follows: Objective weights are calculated based on the dispersion of historical data using the entropy weight method; subjective weights are calculated based on scheduling guidelines and operational requirements using the fuzzy hierarchical analysis method; and a comprehensive weight is obtained through weighted fusion. , The weighting ranges from 0.4 to 0.6. When a scenario of large-scale renewable energy generation is detected, the weighting of the consumption target is increased. During peak grid periods, the weighting of the power fluctuation target is increased. When the SOH of energy storage is below 0.8, the weighting of the life loss target is increased. When carbon prices are high, the weighting of the carbon trading target is increased.

[0048] Furthermore, the specific process of hierarchical distributed solution described in step S5 is as follows:

[0049] 1. Solution of the asymmetric Nash negotiation game at the front-end: Treat each station as an independent interest entity, and define the benchmark payoff for independent operation of entity i as... The actual benefits after collaboration are excess returns Constructing the objective function for asymmetric Nash negotiation:

[0050]

[0051] In the formula, The weighting of negotiating power is calculated by comprehensively considering installed capacity, adjustment flexibility, and historical credit, to meet the following requirements. The augmented Lagrangian function is obtained by using the alternating direction multiplier method for distributed solution:

[0052]

[0053] During the iteration process, each site only uploads the boundary transaction power to the cloud. Core data does not leave the site; the cloud updates dual variables. An internal settlement electricity price is generated to guide each entity to spontaneously adjust its output, eventually converging to the Nash equilibrium solution.

[0054] 2. Intraday Distributed Model Predictive Control: Each site's edge gateway constructs a discrete state-space model of its local subsystem, with a 15-minute rolling period and a 2-hour prediction time domain. It constructs a local objective function that includes tracking deviation penalties, regulation costs, and voltage over-limit penalties. Adjacent sites only exchange boundary power prediction trajectories in the future time domain. They reach consensus on boundary power through iterative consensus protocols and output correction commands on a rolling basis. The CVaR (Conditional Value at Risk) constraint is introduced to transform uncertain scenarios into deterministic robust constraints, with a confidence level set at 0.95.

[0055] 3. Real-time Interpretable MADRL-MPC Hybrid Algorithm: A multi-agent near-end policy optimization algorithm is adopted as the global optimization framework, mapping each energy storage station and charging station as an independent agent. The state space includes the station's real-time operating status, SOC, grid connection point power deviation, and ultra-short-term prediction value; the action space is the charging and discharging power adjustment amount. A channel attention mechanism is introduced into the Critic network, outputting the attention weights of each state feature to achieve decision interpretability. The reference power trajectory output by MADRL is input into the MPC controller to construct a quadratic programming problem.

[0056]

[0057] In the formula, , The reference trajectory output by MADRL. To predict the time domain, The weight matrix is ​​used; the MPC solver outputs the first control command that satisfies all physical hard constraints, realizing a hybrid control of "AI optimization + MPC fallback".

[0058] Compared with the prior art, the present invention has the following significant advantages:

[0059] 1. Cloud-edge collaboration + three-layer time-scale architecture achieves dual improvement in scheduling accuracy and deployment flexibility. This invention deeply integrates three-layer time-scale scheduling with a cloud-edge collaborative architecture. The cloud is responsible for global optimization, while the edge is responsible for rapid response. It fundamentally solves the problem of scheduling target conflicts at different time scales through a "baseline transfer-deviation correction-closed-loop feedback" mechanism, improving the feasibility of scheduling solutions by more than 25%. Furthermore, it reduces communication pressure and single-point failure risk through distributed deployment, adapting to multi-site scenarios with wide-area distribution, and reducing engineering deployment costs by 30%.

[0060] 2. A refined energy storage lifespan model with multi-factor coupling significantly improves overall lifecycle benefits. This invention constructs a two-dimensional degradation model that includes cyclic aging and calendar aging, coupling multiple factors such as depth of discharge, charge / discharge rate, operating temperature, and SOC dwell time range. It accurately quantifies the impact of each charge / discharge decision on battery life, achieving an optimal balance between short-term economic benefits and long-term equipment lifespan. This can extend the lifespan of the energy storage system by 20%-25% and significantly reduce the total lifecycle investment cost.

[0061] 3. Interpretable MADRL-MPC Hybrid Algorithm, Balancing Global Optimization and Safety Compliance. This invention pioneers a hybrid solution architecture that integrates multi-agent reinforcement learning with attention mechanisms and model predictive control. It utilizes MADRL to achieve global economic optimization in high-dimensional nonlinear scenarios, strictly guarantees physical hard constraints through MPC's rolling time-domain optimization, and visualizes key decision-making factors through attention weights. This completely solves the industry pain points of pure AI scheduling, such as "high risk of exceeding limits and lack of interpretability due to its black box nature." The dispatch instruction violation rate is reduced to 0, fully meeting the safety operation requirements of power systems, which are "observable, interpretable, and controllable."

[0062] 4. Asymmetric Nash Negotiation + ADMM Distributed Solution: Achieving a Balance Between Privacy Protection and Fair Interests. This invention deeply integrates asymmetric Nash negotiation game theory with the alternating direction multiplier method. Each power station only needs to exchange coupled variables such as boundary power and settlement price, while core operational data does not leave the station, fully protecting the data privacy of each operator. Simultaneously, based on the installed capacity, adjustment flexibility, and credit rating of each operator, negotiation weights are calculated to achieve a differentiated and fair distribution of excess profits. This significantly enhances the willingness of each operator to cooperate, increasing the renewable energy consumption rate by 8%-10%.

[0063] 5. Multi-scale probabilistic prediction + CVaR robust optimization, balancing economy and safety in extreme scenarios. This invention generates multi-time-scale probabilistic prediction intervals through the STGCN-Transformer model, combined with conditional value at risk robust optimization. This avoids the overly conservative problem of traditional robust optimization and solves the deficiency of insufficient representativeness in stochastic optimization scenarios. It can maintain high scheduling accuracy even under extreme weather conditions, reduce system operation risk by more than 35%, and ensure that the computational efficiency meets the requirements of engineering applications.

[0064] 6. Multi-objective dynamic weighting + low-carbon objective integration enables full-scenario adaptability and policy alignment. This invention constructs a five-dimensional multi-objective optimization system. Through a dynamic weighting adjustment mechanism that integrates subjective and objective factors, it dynamically adjusts the emphasis of each optimization objective according to the operating scenario. It can pursue economic optimization under normal operating conditions, and automatically switch the optimization focus under special scenarios such as large-scale renewable energy generation, grid peak, and equipment aging. At the same time, it incorporates carbon trading costs into the optimization objective, helping the station to participate in carbon market trading and obtain additional revenue, which is in line with the "dual carbon" policy orientation. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the cloud-edge collaborative overall architecture of the wind-solar-storage-charging multi-site collaborative intelligent optimization scheduling system of the present invention;

[0066] Figure 2 This is an overall flowchart of the three-layer multi-timescale collaborative scheduling method of the present invention;

[0067] Figure 3 This is a block diagram of the MADRL-MPC hybrid algorithm that integrates the attention mechanism of this invention.

[0068] Figure 4 This is a diagram of the distributed benefit distribution mechanism based on asymmetric Nash negotiation game in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0070] like Figure 1 As shown, the wind-solar-storage-charging multi-site collaborative scheduling system applicable to this invention adopts a two-tier architecture of "cloud-based scheduling center - site edge gateway," comprising N wind farms, M photovoltaic power generation stations, K electrochemical energy storage stations, and L electric vehicle charging stations connected to the cloud. All stations are connected via the distribution network and ultimately connected to the upper-level power grid. Each station is equipped with an independent edge computing gateway, possessing local data storage, predictive inference, optimization solution, and real-time control functions, and communicates with the cloud-based scheduling center via 5G or a dedicated fiber optic network. The cloud-based scheduling center is responsible for day-ahead global optimization and multi-entity coordination, while each edge gateway is responsible for intraday rolling correction and real-time rapid response.

[0071] Step 1: Cloud-edge collaborative architecture construction and multi-source heterogeneous data preprocessing and multi-scale prediction

[0072] This step establishes the hardware foundation and data support for the entire scheduling method. The specific implementation process is as follows:

[0073] First, the cloud-edge collaborative architecture is deployed: the cloud-based dispatch center is deployed at the regional power grid dispatch master station or virtual power plant operation platform, configured with a high-performance server cluster, running a distributed coordination management module, database, and visualization monitoring system; each charging station's edge computing gateway connects to the station's SCADA system, meteorological monitoring equipment, and charging pile controller to achieve local data acquisition and command issuance. The charging station edge gateway additionally supports V2G reverse power supply control, and can collect user behavior data such as electric vehicle access time, initial SOC, and expected departure time.

[0074] Secondly, multi-source data acquisition and preprocessing were carried out: static parameters of each station were collected through the dispatch data network, including grid connection node location, rated capacity, equipment model, and core operating parameters; dynamic operating data were collected through the SCADA system, including active / reactive power output of wind and solar power, wind speed, irradiance, ambient temperature, energy storage SOC / SOH, charging and discharging power, battery temperature, charging power, and number of online charging piles. The collected data were standardized and preprocessed: outliers were removed using the 3σ criterion, missing data were filled using a combination of linear interpolation and neighbor-value filling, spatiotemporal alignment was achieved by unifying the sampling frequency and timestamp, and the data was mapped to the [0,1] interval through min-max standardization to construct a standardized multi-station operation database.

[0075] Finally, the STGCN-Transformer multi-scale prediction model is constructed: the multi-site system is abstracted as a graph structure, with sites as nodes and electrical distances as edges to build an adjacency matrix. Spatial correlation and short-term temporal features are extracted through gated convolutional layers of a spatiotemporal graph convolutional network, and then input into a Transformer encoder. A multi-head self-attention mechanism is used to capture long-term temporal dependencies such as daily cycles and seasonal alternations. The model outputs three levels of prediction results: deterministic predictions at a 24-hour and 1-hour resolution for the current day, 95% confidence probability intervals at a 4-hour and 15-minute resolution for the intraday day, and rolling predictions at a 15-minute and 1-minute resolution in real time, providing accurate boundary conditions for subsequent hierarchical scheduling.

[0076] Step 2: Construction of a three-layer multi-timescale collaborative scheduling framework

[0077] This invention constructs a three-layer scheduling architecture: "Daily Global Optimization Layer - Intraday Rolling Correction Layer - Real-time Fast Response Layer," with clear division of labor and coordinated operation at each layer, such as... Figure 2 As shown.

[0078] (1) The current global optimization layer

[0079] The day-ahead optimization is performed daily at 24:00 in the cloud, providing global optimization for the next 24 hours with a time resolution of 1 hour, encompassing 24 scheduling periods. Input data includes the next day's weather forecast, time-of-use pricing, charging demand forecast, equipment maintenance plans, grid connection assessment indicators, and carbon emission quotas. The optimization objective is to achieve optimal overall system efficiency through multi-entity collaboration, outputting the next day's 24-hour baseline power plan for each site, the energy storage SOC baseline operating range, and the internal settlement pricing scheme. Constraints include various site operation constraints, grid connection point power upper and lower limits, power fluctuation constraints, system power balance constraints, energy storage initial and final SOC constraints, and carbon emission constraints. The day-ahead optimization results serve as the benchmark for the entire scheduling cycle, propagating down to the edge gateways of each site; intraday scheduling must not significantly deviate from the day-ahead plan boundaries.

[0080] (2) Intraday rolling correction layer

[0081] Intraday rolling correction begins at 00:00 and is executed locally at the edge gateway of each site, rolling every 15 minutes. Each adjustment optimizes the scheduling plan for the next 4 hours, with a time resolution of 15 minutes, for a total of 16 scheduling periods. Input data includes updated ultra-short-term wind and solar forecasts, real-time charging demand, current site operating status, day-ahead plan baseline, and cumulative deviation. The optimization objective is to minimize day-ahead plan deviation while minimizing intraday adjustment costs, correcting power deviation and SOC deviation, and ensuring the feasibility of the day-ahead plan. The rolling mechanism shifts a time window forward every 15 minutes, discarding executed periods and adding new forecast periods for re-optimization. When the cumulative SOC deviation exceeds 5% of the rated capacity, a local correction of the day-ahead plan is triggered, adjusting the baseline values ​​for subsequent periods to prevent the deviation from continuing to expand. The intraday layer reserves 10% of the rated energy storage capacity as a regulation reserve for the real-time layer.

[0082] (3) Real-time fast response layer

[0083] The real-time response layer executes at the edge gateway of each power station every minute, providing rapid optimization for the next 5 minutes, with a time resolution of 1 minute. Input data includes real-time operation data of the power station, real-time power at the grid connection point, second-level wind and solar power fluctuations, and real-time electricity prices. The optimization objective is to quickly smooth out power fluctuations, ensure that the power at the grid connection point meets grid assessment requirements, and maintain system power balance. This is achieved through a millisecond-level rapid decision-making based on an interpretable MADRL-MPC hybrid algorithm, outputting energy storage charging and discharging adjustment commands and orderly charging adjustment commands for charging stations. The real-time adjustment power range must not exceed the reserve capacity range set by the intraday layer, and the State of Charge (SOC) must not exceed the intraday operating range, ensuring that the upper-level scheduling boundaries are not exceeded.

[0084] The three layers form a closed loop through "downward transmission of the benchmark and upward feedback of the deviation," which ensures both the economy of global optimization and the flexibility of local adjustment, fundamentally solving the problem of conflicting scheduling objectives at different time scales.

[0085] Step 3: Modeling the nonlinear decay of the SOH in energy storage with multi-factor coupling

[0086] This step constructs an electrochemical mechanism-based energy storage lifetime decay model to achieve precise quantification and scheduling coupling of lifetime loss.

[0087] First, cyclic aging quantification is performed: the SOC time series within the scheduling period is decomposed cyclically using the rainflow counting method, breaking down irregular SOC fluctuations into several complete charge-discharge cycles and half-cycles. The depth of discharge (DOD), average operating temperature, and charge-discharge rate are recorded for each cycle. A cycle degradation formula is constructed based on the Arrhenius temperature effect equation; the higher the temperature, the faster the internal chemical reaction rate of the battery, and the more significant the capacity degradation. A charge-discharge rate penalty function is introduced; when the charge-discharge rate exceeds the rated rate, the degradation rate increases exponentially, reflecting the damage to the battery caused by high-rate charge-discharge.

[0088] Secondly, calendar aging was quantified: calendar aging is caused by side reactions during battery storage and is related to storage time, ambient temperature, and average SOC level. A square root time model was used to describe the calendar aging pattern, combined with the Arrhenius equation to reflect the effect of temperature, and an SOC influence function was introduced to reflect the accelerating effect of high SOC storage on calendar aging.

[0089] Finally, total capacity decay and cost reduction are calculated: the capacity decay of cyclic aging and calendar aging are added together to obtain the total capacity decay within the scheduling cycle, which is then divided by the rated capacity to obtain the SOH decay rate. The SOH decay is converted into an equivalent economic cost, proportionally converted according to the total lifecycle investment cost, and incorporated into the optimization objective function, so that the scheduling algorithm automatically balances short-term gains with long-term lifecycle losses when making decisions.

[0090] Step 4: Construction of a multi-objective dynamic weight optimization model

[0091] This invention constructs a five-dimensional multi-objective optimization system and realizes dynamic adaptive adjustment of weights.

[0092] The overall system optimization objective is a weighted sum of five sub-objectives, each corresponding to one of the five dimensions: economic efficiency, grid absorption capacity, equipment lifespan, grid security, and user experience. To avoid the problem of fixed weights failing to adapt to complex scenarios, a dynamic weighting mechanism that integrates objective and subjective factors is adopted: objective weights are calculated using the entropy weighting method, based on historical operating data from the past 30 days. Indicators with lower information entropy have greater dispersion and more information, and therefore higher weights, making it entirely data-driven; subjective weights are calculated using fuzzy hierarchical analysis, constructing a judgment matrix based on grid dispatching guidelines, system operating requirements, and scenario priorities, reflecting the dispatching principle of prioritizing safety. The combined weight is obtained by weighting and fusing these two factors.

[0093] Based on this, a dynamic adjustment mechanism is set up to trigger specific scenarios: when a surge in renewable energy generation is predicted and the curtailment rate is expected to exceed 5%, the weight of the renewable energy consumption target is increased by 0.1 to 0.2, while the weight of the operating cost target is correspondingly reduced; when the grid is in peak electricity consumption, the weight of the power fluctuation target at the grid connection point is increased to ensure grid stability; when the state of harmlessness (SOH) of any energy storage station is below 80%, the weight of the lifespan loss target is increased to prioritize the protection of aging equipment; when carbon trading prices are high, the weight of the carbon trading cost target is increased to enhance the effectiveness of low-carbon dispatching. Through dynamic weight adjustments, the optimized model can adapt to different operating scenarios and always maintain the globally optimal dispatching strategy.

[0094] Step 5: Hierarchical Distributed Intelligent Optimization Solution

[0095] Differentiated solution algorithms are adopted to achieve a balance between efficiency and accuracy for the different time scale characteristics of the three-layer scheduling architecture.

[0096] (1) Solving the asymmetric Nash negotiation game at the front-end + ADMM

[0097] In recent times, for multi-party collaborative scenarios, distributed solution has been adopted to protect data privacy and achieve fair distribution of benefits, such as... Figure 4 As shown, the independent operating baseline returns of each entity are first defined, i.e., the optimal returns when not participating in coordination, as the negotiation breakdown point. An asymmetric Nash negotiation objective is constructed to maximize the weighted product of the excess returns of each entity. The negotiation weights comprehensively consider installed capacity, adjustment flexibility, and historical credit, reflecting the fairness principle of "the greater the contribution, the greater the return".

[0098] A distributed iterative solution is employed using the alternating direction multiplier method: Global coupling and dual variables are initialized in the cloud and broadcast to each power station; each power station's edge gateway constructs a sub-optimization model locally, aiming to maximize its own overall benefit, incorporating the deviation from the global variables as a penalty term into the objective function, independently solving for local switching power and uploading it to the cloud; after collecting the switching power of all power stations, the cloud updates the global coupling and dual variables, calculates the original residual and dual residual, and determines whether convergence has occurred. During the iteration process, each power station only uploads boundary trading power; core cost parameters, equipment parameters, and other private data are completely retained locally, fully ensuring data security. After convergence, the day-ahead switching power plan and internal settlement price for each power station are obtained, serving as the benchmark for intraday scheduling.

[0099] (2) Intraday Distributed Model Predictive Control

[0100] The intraday layer employs distributed model predictive control to achieve rolling correction. Each site's edge gateway constructs a discrete state-space model of its local subsystem. State variables include energy storage SOC, equivalent SOC of electric vehicle virtual energy storage, and reactive power output, while control variables include energy storage charging and discharging power adjustment and charging station charging power adjustment.

[0101] In each rolling cycle, each power station constructs a finite-time-domain optimization objective based on the latest ultra-short-term forecast data, including deviation penalties for tracking day-ahead plans, regulation cost penalties, and voltage limit exceedance penalties. Adjacent power stations communicate via edge gateways, exchanging boundary power trajectories within the future forecast time domain. Through iterative updates using a consensus protocol, all power stations' boundary power predictions converge. A CVaR (Conditional Value at Risk) constraint is introduced to transform uncertain scenarios into deterministic constraints, ensuring operational safety with a 95% probability. After convergence, the control actions for the current moment are extracted and executed, progressing continuously over the rolling time window.

[0102] (3) Real-time layer interpretability MADRL-MPC hybrid algorithm

[0103] For real-time scenarios with second-level fluctuations, a hybrid intelligent algorithm is employed to balance speed, security, and interpretability, such as... Figure 3 As shown.

[0104] A multi-agent near-end policy optimization algorithm is adopted as the basic framework, with one agent corresponding to each energy storage station and charging station. The state space includes features such as local real-time power output, energy storage SOC, grid connection point power deviation, and ultra-short-term prediction deviation; the action space is the continuous adjustment of charging and discharging power. A channel attention module is introduced into the Critic network, which automatically learns the importance weights of each state feature through compression-excitation operations. The attention weights are displayed synchronously when the decision is output, allowing operators to clearly identify the core driving factors of the decision and achieving interpretability.

[0105] To address the issue of pure reinforcement learning being prone to exceeding limits, MPC is introduced as a safety fallback module. The reference power trajectory output by MADRL is used as the tracking target of MPC. Based on a local accurate physical model, MPC constructs a quadratic programming problem, rigorously considering all physical constraints such as charging and discharging power limits, SOC range, ramp rate, and node voltage, to solve for the optimal control command that satisfies all constraints. Only the first step of the control is issued and executed; the new system state is fed back to MADRL's experience replay pool for model iterative optimization. This hybrid architecture retains the global optimization capability and fast response speed of AI while completely eliminating the risk of exceeding limits through MPC, achieving a safe implementation of intelligent scheduling.

[0106] Step 6: Issuance of scheduling instructions and closed-loop feedback of operating status

[0107] After optimization calculations are completed, dispatch instructions are issued in a tiered manner: the daily plan is issued to the operators of each site through the dispatch data network as a reference for the next day's operation; intraday correction instructions are issued to the site control system for automatic execution through the AGC / AVC system; and real-time adjustment instructions are issued to the energy storage PCS and charging stack management system through the fast communication channel to achieve millisecond-level response.

[0108] The dispatch center and edge gateway collect real-time operational data from each site, calculating the deviation between actual and planned values, including power deviation, SOC deviation, and energy deviation. When the deviation exceeds a set threshold, it is immediately fed back to the intraday rolling correction layer and real-time response layer for correction in the next dispatch cycle. After the daily operation ends, the dispatch effect of the day is evaluated, and indicators such as renewable energy absorption rate, operating cost, lifespan loss, and carbon emissions are calculated. The evaluation results are used for weight adjustment and model parameter optimization.

[0109] An online incremental learning mechanism is established: the edge gateway continuously calculates the average reward for real-time scheduling. If the reward consistently falls below a preset threshold, the reinforcement learning policy network is fine-tuned using recent operational data with small samples to adapt to seasonal changes in wind and solar power output and charging behavior. The prediction model is also regularly updated based on the latest data to continuously improve prediction accuracy. This ultimately forms a complete closed loop of "decision-execution-feedback-optimization".

[0110] Example Effect Verification

[0111] To verify the effectiveness of the method of the present invention, a simulation was conducted using a system in a certain region that includes 2 wind farms (total capacity 100MW), 3 photovoltaic power stations (total capacity 150MW), 2 energy storage stations (total capacity 50MW / 100MWh), and 4 charging stations (total power 20MW) as an example. The operational performance of the method of the present invention was compared with that of the traditional centralized single-objective scheduling method and the pure reinforcement learning scheduling method. The results are as follows:

[0112] 1. Renewable energy consumption level: The consumption rate of traditional methods is 89.2%, that of pure DRL method is 93.5%, and that of the method of this invention is 97.8%, which is 8.6 percentage points higher than that of traditional methods, and the amount of wind and solar power curtailed is reduced by 72%;

[0113] 2. Overall system operating cost: The traditional method costs approximately RMB 286,000 per day, the pure DRL method costs approximately RMB 261,000 per day, and the method of this invention costs approximately RMB 241,000 per day, which is 15.7% lower than the traditional method;

[0114] 3. Energy storage lifespan loss: The traditional method has a daily SOH decay of 0.012%, the pure DRL method has a decay of 0.011%, and the method of this invention has a decay of 0.0091%, which is a 24.2% reduction in decay rate and a corresponding lifespan extension of about 32%.

[0115] 4. Power grid operation safety: The traditional method has a power fluctuation rate of 4.2% at the grid connection point and 3 power limit violations under extreme weather conditions; the pure DRL method has a fluctuation rate of 2.8% and 5 command limit violations; the method of this invention has a fluctuation rate of 1.8%, a command limit violation rate of 0%, and zero limit violations under extreme scenarios.

[0116] 5. Multi-entity benefits: After adopting the game-theoretic distribution mechanism of this invention, energy storage stations obtain 38% of excess benefits due to their high adjustment flexibility, wind and solar power stations obtain stable benefits by reducing curtailment, and charging stations reduce electricity costs while ensuring user satisfaction. All entities obtain positive excess benefits, and the willingness to cooperate is significantly improved.

[0117] In summary, the method of this invention is significantly superior to existing technologies in multiple dimensions, including economy, absorption capacity, equipment lifespan, grid security, and multi-entity collaboration, and possesses outstanding technical advantages and engineering application value.

[0118] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations, characterized in that: Includes the following steps: S1. Construct a two-level cloud-edge collaborative architecture of "cloud dispatch center - site edge gateway" to perform topology identification and parameter collection for wind farms, photovoltaic power generation stations, electrochemical energy storage stations, and electric vehicle charging stations in the region. Perform cleaning, normalization, and spatiotemporal alignment processing on multi-source heterogeneous operation data. Construct the STGCN-Transformer hybrid prediction model to generate probability prediction intervals for wind and solar power output and charging load at three time scales: day-ahead, intraday, and real-time, providing data support for hierarchical dispatch. S2. Build a three-layer scheduling architecture consisting of "day-ahead global optimization layer - intraday rolling correction layer - real-time rapid response layer", clarify the time resolution, optimization objectives, decision variables and constraint boundaries of each layer, and establish a benchmark transfer and deviation feedback mechanism between upper and lower layers; the day-ahead global optimization layer is deployed in the cloud scheduling center, and the intraday rolling correction layer and real-time rapid response layer are deployed in the field edge gateway. S3. Differentiate between the dual degradation mechanisms of cycle aging and calendar aging, use the rainflow counting method to decompose the charge-discharge cycle, combine the Arrhenius temperature effect equation and the charge-discharge rate penalty function to construct a multi-factor coupled energy storage battery health state degradation model, quantify the capacity degradation caused by a single charge-discharge process, and convert it into life loss cost and incorporate it into the optimization target. S4. Establish a five-dimensional multi-objective optimization function with the core objectives of minimizing the overall system operating cost, maximizing the renewable energy absorption rate, minimizing energy storage life loss, minimizing power fluctuation at the grid connection point, and minimizing carbon trading costs. Construct a dynamic weight adjustment mechanism that integrates subjective and objective factors based on the entropy weight method and fuzzy hierarchical analysis method, and automatically adapt the weight coefficients of each optimization objective according to the real-time operating scenario. S5. Differentiated solution algorithms are adopted for the different characteristics of the three-layer scheduling architecture: the day-ahead layer is based on an asymmetric Nash negotiation game to construct a distributed collaborative optimization model and uses the alternating direction multiplier method for solution; The intraday layer employs distributed model predictive control for rolling corrections to track daily plan deviations; the real-time layer uses a MADRL-MPC hybrid algorithm that incorporates an attention mechanism to output fast decision instructions while ensuring constraint compliance and decision interpretability. S6. Decompose and distribute the scheduling instructions generated at each level to the corresponding stations for execution, collect the operating data of each station in real time and calculate the deviation between the actual operating status and the planned value, and feed the deviation back to the intraday rolling correction layer and the real-time response layer; establish an online incremental learning mechanism to continuously optimize the prediction model and scheduling algorithm, and realize closed-loop iterative optimization of the scheduling process.

2. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S1, the STGCN-Transformer hybrid prediction model is constructed by abstracting the multi-site system into a graph structure. Where V is the set of station nodes, E is the set of edges, and A is the adjacency matrix reflecting the electrical distance and geographical correlation between stations; firstly, spatial topological features and short-term temporal dependencies are extracted through a spatiotemporal graph convolutional network, and the calculation formula for the gated temporal convolutional layer is: In the formula, The input features are for the l-th layer. This represents the graph convolution operation. , For learnable parameters, The Sigmoid activation function is used; the spatial feature sequence is then input into the Transformer encoder to capture long-distance time series dependencies; finally, a deterministic prediction benchmark for the previous 24 hours, a 95% confidence probability interval for the intraday 4 hours, and real-time ultra-short-term rolling prediction values ​​are generated.

3. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S2, the specific configuration of the three-layer scheduling architecture is as follows: Currently, the global optimization layer has a time resolution of 1 hour and an optimization cycle of 24 hours, outputting the baseline power plan and energy storage SOC baseline range for each site as hard constraints for the lower layer. The intraday rolling correction layer has a time resolution of 15 minutes, is executed every 15 minutes, has an optimization cycle of 4 hours, corrects the deviation between power and SOC, and reserves 10% of the rated power for adjustment. The real-time fast response layer has a time resolution of 1 minute, executes every 1 minute, and has an optimization cycle of 5 minutes, which is used to smooth out real-time power fluctuations. The constraint transmission mechanism between upper and lower layers is as follows: the power plan output by the upper layer and the SOC range serve as hard constraints for the lower layer, and the optimization of the lower layer must not exceed the boundary of the upper layer; when the cumulative SOC deviation exceeds 5% of the rated capacity, the rolling correction of the upper layer plan is triggered.

4. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S3, the process of constructing the energy storage battery health state degradation model is as follows: (1) Cyclic aging calculation: The SOC fluctuation sequence is decomposed into N charge-discharge cycles using the rainflow counting method, and the discharge depth of the i-th cycle is extracted. Average operating temperature With charge / discharge rate The capacity decay per cycle is: In the formula, These are the fitting coefficients for the battery materials. The activation energy is given by R, where R is the ideal gas constant. This is the multiplier penalty function; (2) Calendar aging calculation: The calculation formula is as follows In the formula, Calendar aging factor. For calendar aging activation energy, The average temperature over a period of time. Let SOC stay influence function; (3) Total attenuation and cost conversion: The total SOH attenuation is the sum of the attenuation due to cycle aging and calendar aging divided by the rated capacity. The life loss cost is the converted value of the initial investment cost and the attenuation.

5. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S4, the specific form of the multi-objective optimization function is as follows: in, The overall operating cost of the system includes the cost of purchasing and selling electricity from the grid, the cost of energy storage operation and maintenance, the cost of penalties for curtailing wind and solar power, and the cost of carbon trading. For the curtailment rate of renewable energy; Energy storage lifespan loss rate; Power fluctuation rate at the grid connection point; This results in a loss of user satisfaction with charging services.

6. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 5, characterized in that: In step S4, the dynamic weight adjustment mechanism is as follows: Objective weights are calculated based on the dispersion of historical data using the entropy weight method; subjective weights are calculated based on scheduling and operation requirements using the fuzzy hierarchical analysis method; and the weighted fusion is used to obtain the comprehensive weight. , The value ranges from 0.4 to 0.6; when scenarios such as a surge in new energy generation, peak grid activity, energy storage SOH below 0.8, and high carbon prices are detected, the weight of the relevant optimization targets is increased accordingly.

7. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S5, the specific solution process for the day-ahead layer is as follows: Each station is considered an independent interest entity, and the independent operating benchmark benefit of entity i is defined as... The actual benefits after collaboration are excess returns Constructing the objective function for asymmetric Nash negotiation: In the formula, The negotiating power weight is calculated by combining installed capacity, adjustment flexibility, and historical credit. The distributed solution is obtained by using the alternating direction multiplier method. During the iteration process, each power station only uploads the boundary trading power, and the cloud generates the internal settlement electricity price by updating the dual variables, which converges to the Nash equilibrium solution.

8. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S5, the specific solution process for the intraday layer is as follows: Each site edge gateway constructs a discrete state-space model of its local subsystem, with a rolling period of 15 minutes and a prediction time domain of 2 hours, and constructs a local objective function that includes tracking deviation penalty, adjustment cost and voltage over-limit penalty; adjacent sites exchange boundary power prediction trajectories and reach consensus through consensus protocol iteration; introduce CVaR conditional value of risk constraint with a confidence level of 0.95 to transform the uncertain scenario into a deterministic robust constraint, and output correction instructions on a rolling basis.

9. The intelligent optimization scheduling method for multi-site collaborative operation of wind, solar, storage, and charging facilities according to claim 1, characterized in that: In step S5, the specific solution process of the real-time layer is as follows: a multi-agent near-end policy optimization algorithm is adopted as the framework to map energy storage stations and charging stations as independent agents; a channel attention mechanism is introduced into the Critic network to output the attention weights of the state features to achieve decision interpretability; the reference power trajectory output by MADRL is input into the MPC controller to construct a quadratic programming problem with hard constraints of the equipment, and output control commands that satisfy the physical constraints to achieve hybrid control of "AI optimization + MPC fallback".

10. The method for collaborative intelligent optimization scheduling of multiple wind, solar, storage, and charging stations according to claim 1, characterized in that: In step S6, scheduling instructions are issued in a tiered manner: daily plans are issued through the scheduling data network, intraday correction instructions are automatically executed through the AGC / AVC system, and real-time adjustment instructions achieve millisecond-level response through a fast communication channel; when the running deviation exceeds the set threshold, the lower-level correction is triggered, and the scheduling effect is evaluated and the model parameters are optimized after the daily operation is completed; the online incremental learning mechanism is as follows: when the real-time scheduling average reward is continuously lower than the threshold, the reinforcement learning model is fine-tuned with small samples using recent data to continuously adapt to changes in operating conditions.

Images