A method and system for intelligent scheduling and load distribution in a multi-user green power direct connection scenario

Abstract

The application discloses a kind of intelligent scheduling and load distribution method and system under multi-user green electricity direct connection scene.The method collects new energy generation, user load, energy storage, power market and carbon emission and pre-processes multiple-source heterogeneous data;K-means weather typing cluster is used to construct source-load prediction model in combination with XGBoost regression, and the prediction accuracy is improved by sliding window residual online correction;Under the condition of meeting user fairness constraint and individual constraint, a multi-objective dynamic allocation model is established to maximize green power consumption, minimize electricity cost and minimize carbon emission;Improved NSGA-III and multi-agent reinforcement learning MAPPO are used for hybrid optimization solution, and closed-loop rolling correction is realized through model predictive control;The allocation result is uploaded to the alliance blockchain, and the green electricity cost settlement and traceability are completed through smart contract.The application realizes the fair, economic, low-carbon distribution and credible traceability of multi-user green electricity.

Description

Technical Field

[0001] This invention belongs to the field of new energy power system dispatching technology, specifically relating to a multi-user intelligent dispatching and load allocation method and system in a green power direct connection scenario. Background Technology

[0002] With the large-scale integration of new energy sources into the power system, the green electricity direct connection model is expanding from a single-user "point-to-point" model to a multi-user "clustered sharing" model. The "Notice on Relevant Matters Concerning the Orderly Promotion of the Development of Multi-User Green Electricity Direct Connection" (NDRC Energy

[2026] No. 688), jointly issued by the National Development and Reform Commission and the National Energy Administration, clarifies that multi-user green electricity direct connection can be carried out in scenarios such as industrial parks, zero-carbon parks, and computing facilities. It also requires that each user within the project allocate their self-generated and self-consumed green electricity according to the proportion of electricity consumption in each time period, so as to achieve precise matching of hourly new energy generation and consumption.

[0003] In the actual operation of multi-user green power direct connection, new energy power generation (wind power, photovoltaic) has inherent intermittency and volatility, while the load characteristics of multiple users are different, and policies impose strict constraints on source-load matching: the proportion of self-generated and self-consumed electricity in the total available power generation shall not be less than 60%, and the proportion of self-generated and self-consumed electricity in the total available power generation shall not be less than 30%. This brings severe challenges to intelligent scheduling and load allocation.

[0004] The existing technology has the following shortcomings: (1) The dispatch scheme lacks a comprehensive optimization mechanism for the dynamic priority of multiple users and policy constraints. Existing dispatch methods mostly target a single economic indicator and do not comprehensively consider multiple factors such as the urgency of users' carbon emission reduction, electricity price sensitivity, and load adjustability, making it difficult to meet the coordinated requirements of the differentiated needs of multiple users and the policy constraints of direct green electricity connection.

[0005] (2) There is a lack of predictive scheduling capability for source load uncertainty, and the prediction methods are relatively conventional. Existing prediction techniques mostly use CNN-LSTM or simple Seq2Seq models, which fail to make full use of weather classification information and real-time residual correction, resulting in limited prediction accuracy and thus affecting the scheduling optimization effect.

[0006] (3) Insufficient fairness in allocation among multiple users and insufficient accuracy in green electricity traceability. Document No. 688 requires "allocating self-generated and self-consumed green electricity according to the proportion of electricity consumption in each time period to achieve hourly precise matching", but existing technologies mostly remain at the level of rough allocation and cannot achieve refined and efficient time-varying load allocation and green electricity traceability.

[0007] (4) The electricity-carbon synergy mechanism has not yet been effectively embedded in the scheduling system. The low-carbon attributes of green electricity need to be "what you see is what you get" at the scheduling layer, but existing scheduling models mostly treat the electricity-carbon attributes as offline statistical indicators rather than online optimization targets, which cannot achieve dynamic optimal scheduling of electricity-carbon synergy.

[0008] Therefore, there is an urgent need for an intelligent scheduling method that can comprehensively predict power output, flexibly adjust load, optimize energy storage, and dynamically allocate multiple objectives for multi-user green electricity direct connection scenarios. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention proposes an intelligent scheduling and load allocation method and system for multi-user green electricity direct connection scenarios. The core of this invention lies in employing weather-based clustering and XGBoost regression for source-load prediction, and introducing online correction of sliding window residuals; constructing a multi-objective allocation model driven by user fairness and personalization constraints; adopting a hybrid solution strategy combining the improved non-dominated sorting genetic algorithm (NSGA-III) and multi-agent proximal policy optimization (MAPPO); and integrating a blockchain-based green electricity traceability mechanism to form a complete technical closed loop of "prediction-allocation-correction-proofing".

[0010] The specific technical solution is as follows: A method for intelligent scheduling and load allocation in a multi-user green electricity direct connection scenario includes the following steps: S1. Collecting multi-source heterogeneous data, including new energy power generation, user load, energy storage system, electricity market and carbon emission data, and preprocessing them; S2. Performing weather type clustering on historical meteorological data, constructing new energy output prediction models and user load prediction models based on the weather type results, and obtaining source-load prediction values ​​through online correction of sliding window residuals; S3. Based on the source-load prediction values, under the conditions of satisfying user fairness constraints and user personalization constraints, establishing a multi-objective dynamic allocation model with the objectives of maximizing green electricity consumption rate, minimizing electricity costs and minimizing carbon emissions; S4. Using evolutionary computation and multi-agent reinforcement learning for hybrid optimization solution to obtain the green electricity allocation scheme for each user; S5. Monitoring the deviation of the green electricity allocation scheme in real time, and using model predictive control for closed-loop rolling correction when preset trigger conditions are met; S6. Uploading the corrected allocation results and electricity consumption data to the consortium blockchain, and using smart contracts for green electricity fee settlement and green electricity traceability and evidence storage.

[0011] Furthermore, in step S1, the new energy power generation data includes photovoltaic output, wind speed and direction, irradiance, ambient temperature and photovoltaic module backsheet temperature; the user load data includes each user's electricity load curve, demand response willingness and adjustable load capacity, and production plan information; the energy storage system data includes energy storage status of charge, charging and discharging power, health status, and temperature; the electricity market data includes real-time electricity price, medium- and long-term contract electricity price, and spot market electricity price forecast; and the carbon emission data includes each user's industry carbon quota, real-time grid carbon emission factor, and carbon trading market price.

[0012] Furthermore, in step S2, the weather type clustering uses the K-means algorithm to divide historical meteorological data into different weather types, each corresponding to different renewable energy output characteristics and user load patterns; the renewable energy output prediction includes: constructing a feature vector X(t) containing meteorological features, time features, historical statistical features, and weather type unique thermal coding features, training an XGBoost regressor, using X(t) as input, and outputting a basic predicted value of renewable energy output, employing a squared loss function and an early shutdown strategy; the multi-user load prediction includes: independently training an XGBoost regressor for each user, whose input features include meteorological features, time features, the user's historical load sequence, and industry electricity consumption characteristic indicators, the industry electricity consumption characteristic indicators including at least unit output value electricity consumption, interruptible load ratio, and adjustable load ratio.

[0013] Furthermore, in step S2, the online correction of the sliding window residual specifically involves: calculating the average prediction error over the past L time periods as the correction amount, weighting it with an adaptive learning rate, and then adding it to the source load base prediction value; the correction amount is updated continuously with real-time operating data to dynamically track the actual output and load trends, thereby obtaining the hourly corrected prediction value.

[0014] Furthermore, in step S3, the constraints of the multi-objective allocation optimization model also include source-load balance constraints, policy-based electricity allocation constraints, and energy storage operation constraints; wherein the policy-based electricity allocation constraints include lower limits for self-consumption ratios, lower limits for green electricity coverage ratios, and upper limits for grid-connected electricity ratios; and the energy storage operation constraints include energy storage state-of-charge operation range constraints and energy storage state-of-charge time-series recursive constraints.

[0015] Furthermore, in step S3, the user fairness constraint uses the Gini coefficient to quantify the fairness of each user's green electricity allocation share. The green electricity allocation share is the proportion of green electricity allocated to each user to the user's total load, and the Gini coefficient does not exceed a preset maximum allowable threshold. The user-personalized constraint includes the user's minimum green electricity utilization rate constraint and the user's maximum affordable electricity price constraint. The constraint parameters can be set by each user and transmitted to the allocation optimization model in real time to participate in the scheduling calculation.

[0016] Furthermore, in step S3, the green electricity consumption maximization target takes maximizing the ratio of the total green electricity power allocated to each user in each time period within the scheduling cycle to the total predicted value of new energy output correction as the optimization direction, which is used to increase the self-consumption ratio of green electricity direct connection projects; the electricity cost minimization target, when solving for energy costs, uses the sum of the fixed electricity cost of green electricity direct connection and the dynamic grid purchase cost as the user-side energy cost, and the dynamic grid purchase cost is determined based on the real-time changing comprehensive grid electricity price and the net power purchase amount determined by the portion of the load prediction value exceeding the green electricity allocated to the user; and, on the energy storage cost side, the electricity cost minimization target simultaneously introduces the dynamic factor affected by the charging and discharging power amplitude. The system incorporates the lifespan degradation cost and fixed operation and maintenance costs measured by operating period to form a real-time dynamic cost allocation mechanism for the entire life cycle of energy storage. When calculating carbon emissions, the carbon emission minimization objective uses the real-time dynamically updated grid average carbon emission factor and the net purchased power to calculate the total implicit carbon emissions generated by the system's purchase of electricity from the external grid throughout the entire time period, and minimizes this total amount as the optimization direction. The green electricity absorption rate maximization objective, the electricity cost minimization objective, and the carbon emission minimization objective together serve as parallel optimization objectives for multi-objective optimization. These objectives are used in the day-ahead or real-time scheduling of green electricity direct-connection systems to collaboratively optimize the green absorption level, economy, and low-carbon performance of the system operation under the premise of ensuring power balance constraints.

[0017] Furthermore, in step S4, the evolutionary computation employs an improved non-dominated sorting genetic algorithm NSGA-III to search for the global Pareto optimal solution set for the current day. The improvements include: using Latin hypercube sampling combined with a uniform design table to generate the initial population, ensuring a uniform distribution of the initial solutions in the decision space; after each preset number of generations, using K-means clustering to regenerate adaptive reference points based on the actual distribution of the current Pareto front, so that the reference points match the shape of the front; introducing a constraint dominance priority strategy, sorting infeasible solutions according to constraint violation values, with feasible solutions taking precedence over infeasible solutions; after each generation of evolution, using a differential evolution operator to perform a local search on the individuals at the Pareto front; and NSGA-III outputs a set of current day scheduling schemes as the benchmark scheme for subsequent intraday rolling corrections.

[0018] Furthermore, in step S4, the multi-agent reinforcement learning uses the multi-agent near-end policy optimization (MAPPO) algorithm to perform intraday real-time dynamic correction of the day-ahead scheduling scheme. This includes: constructing four cooperative agents: power agent, energy storage agent, distribution agent, and balancing agent; adopting a centralized training and distributed execution architecture; where the observation space and action space of each agent correspond to the state and control variables of the managed objects, respectively; and setting a cooperative reward item in the global reward function, which is positively correlated with the degree of system power balance under the joint action of all agents.

[0019] Furthermore, in step S5, the preset triggering conditions include one of the following five categories: the relative deviation between the actual output of new energy and the intraday rolling optimization benchmark value exceeds a first percentage threshold; the relative deviation between the actual total load of multiple users and the intraday rolling optimization benchmark value exceeds a second percentage threshold; the energy storage state of charge exceeds a preset allowable operating range; the fluctuation of the real-time electricity price compared to the predicted value exceeds a third percentage threshold; the change in user load within a preset time window exceeds a preset sudden change threshold or there is a fault in the electrical equipment; the closed-loop rolling correction adopts a model predictive control method. In each control cycle, with the current operating state as the initial point, within a preset prediction time domain with a preset resolution as the step size, minimizing the cumulative power deviation within the prediction time domain is the optimization objective. The control scheme obtained by intraday rolling optimization is used as the reference benchmark. The multi-agent reinforcement learning is called again to solve for the optimal control sequence. Only the correction control command at the current moment is executed. In the next control cycle, the above process is repeated according to the updated operating state to achieve a continuous closed loop of day-ahead scheduling, intraday rolling, and real-time correction.

[0020] Furthermore, in step S6, the consortium blockchain adopts a Hyperledger Fabric or FISCO BCOS architecture, and the smart contract includes: recording the green electricity allocation and actual electricity consumption of each user in each time period on the blockchain for evidence storage to meet the hourly green electricity traceability requirements; automatically executing green electricity fee settlement according to preset green electricity settlement rules; generating hourly resolution green electricity traceability proof for each user, and providing trusted data credentials for carbon emission accounting and green certificate issuance.

[0021] A smart scheduling and load allocation system for multi-user green electricity direct connection scenarios, used to execute the method, includes: a data acquisition layer for acquiring multi-source heterogeneous data and performing preprocessing; a prediction layer for performing weather type clustering, source-load prediction model construction, and online correction of sliding window residuals; an optimization decision layer for establishing the multi-objective dynamic allocation model and performing hybrid optimization solution; a real-time control layer for performing real-time deviation monitoring and closed-loop rolling correction; a user interaction interface layer for receiving personalized parameters set by users and transmitting them to the optimization decision layer in real time as personalized constraints for users, and displaying green electricity allocation results and carbon emission reports to each user; and a blockchain evidence storage layer for performing green electricity fee settlement and green electricity traceability evidence storage, which is deployed with smart contracts and automatically performs green electricity fee settlement and generates hourly green electricity traceability certificates when trigger conditions are met.

[0022] Beneficial effects (1) The present invention adopts a source load prediction method combining weather classification and XGBoost. By dividing the weather type through K-means clustering, XGBoost regressors are trained to predict the source load. This avoids the defects of complex training and poor interpretability of deep learning. While ensuring accuracy, the calculation time is shortened to the second level, providing a highly reliable prediction boundary for day-ahead scheduling.

[0023] (2) The present invention introduces a sliding window residual online correction mechanism. The correction amount is calculated by the prediction error of the historical period and then superimposed on the basic prediction value after being weighted by the adaptive learning rate. The correction amount is updated with the real-time operation data, dynamically tracking the actual output and load trend, effectively reducing the impact of prediction deviation on scheduling optimization.

[0024] (3) This invention constructs a multi-objective dynamic allocation model with the parallel optimization objectives of maximizing green electricity consumption rate, minimizing electricity cost and minimizing carbon emissions. It breaks through the limitations of a single economic indicator and coordinates the optimization of the green consumption level, economy and low carbon performance of the system operation under the premise of ensuring power balance.

[0025] (4) This invention incorporates policy matching constraints and energy storage state of charge constraints, and uses the Gini coefficient to quantify the fairness of the green electricity allocation share of each user. Combined with personalized parameters such as the minimum green electricity utilization rate constraint and the maximum affordable electricity price constraint, it meets the synergistic requirements of the differentiated needs of multiple users and the policy constraints of direct green electricity connection, and achieves the joint optimization of policy compliance, economy, greenness and fairness.

[0026] (5) The present invention designs a day-day and intraday dual-layer optimization architecture. The day-day layer uses an improved NSGA-III to solve multi-objective problems, and the intraday layer uses a multi-agent MAPPO to correct deviations in real time. A cooperative reward item is introduced to coordinate four types of agents, deeply coupling energy storage and distribution to achieve the synergy of global optimization and rapid response.

[0027] (6) This invention combines model predictive control with multi-agent reinforcement learning to monitor power, load, energy storage charge status and line status at high speed. When the deviation exceeds the standard, it triggers closed-loop rolling correction, which can respond to power fluctuations and emergency conditions in milliseconds, avoid the risk of allocation failure or exceeding limits caused by prediction deviation, and significantly improve the system safety and robustness.

[0028] (7) This invention introduces a consortium blockchain to provide tamper-proof evidence of the allocation results and electricity consumption data for each time period. Through smart contracts, it automatically generates hourly green electricity traceability certificates, allocates green certificates, and calculates carbon emission reports, solving the problems of low credibility of green electricity records and difficulty in electricity-carbon coordination under the multi-user mode, and realizing automatic settlement of green electricity fees. Attached Figure Description

[0029] Figure 1This is a flowchart of an intelligent scheduling and load allocation method for a multi-user green electricity direct connection scenario according to the present invention; Figure 2 Weather classification and XGBoost forecast flowchart; Figure 3 Multi-agent Proximal Policy Optimization (MAPPO) collaborative framework diagram. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention are further illustrated below with reference to the accompanying drawings and product effect diagrams.

[0031] Example 1: A method for intelligent scheduling and load distribution of green electricity directly connected to multiple users, such as... Figure 1 : Step S1: Data Acquisition and Preprocessing Through the data acquisition layer deployed in the multi-user green electricity direct connection system, data is collected and recorded in real time, specifically including: New energy side data: hourly actual output power of photovoltaic array, wind speed and direction, irradiance, ambient temperature, photovoltaic module backsheet temperature, etc. User-side data: electricity load curves of each user, willingness to participate in demand response and adjustable load capacity, production plan information; Energy storage system data: State of charge (SOC), charging and discharging power, health status, and temperature of the energy storage power station; Electricity market data: real-time electricity prices, medium- and long-term contract electricity prices, and spot market electricity price forecasts; Carbon emission data: carbon allowances for each user's industry, real-time grid carbon emission factors, and carbon trading market prices.

[0032] After performing outlier detection (using the isolated forest algorithm) and missing value imputation (using time-series spline interpolation) on the above multi-source heterogeneous data, Z-score normalization is performed to eliminate the influence of dimensions.

[0033] Step S2: Source-load co-prediction based on weather morphology and XGBoost This step employs a combined prediction method of weather typology clustering, XGBoost regression, and sliding window residual correction, avoiding complex deep learning models and achieving a balance of high accuracy, interpretability, and computational efficiency. Figure 2 .

[0034] (1) Historical weather classification Historical meteorological data (irradiance, temperature, humidity, wind speed, cloud cover) were collected, and the K-means algorithm was used to classify weather types into five categories: clear skies, partly cloudy, cloudy, overcast / rainy, and strong winds. Each weather category corresponds to different renewable energy output characteristics and user load patterns. The classification results are used for sample grouping during subsequent model training to improve prediction accuracy.

[0035] (2) Feature engineering construction For the time period t to be predicted, construct the feature vector X(t), including: Meteorological characteristics: air temperature T (t), irradiance G (t), wind speed V (t), humidity H (t); Time characteristics: hour (t), day of the week (t), holiday (t); Historical characteristics: Average output of new energy sources in the first three periods Output value at the same time the previous day ; Classification characteristics: the unique thermal code of the current weather category. onehot (t).

[0036] (3) XGBoost prediction model New energy output prediction: Training the XGBoost regressor Input X(t), output predicted value

[0037]

[0038] in This is a basic forecast value for new energy processing. For the XGBoost model trained on historical data, a squared loss function and an early stopping strategy are employed.

[0039] Multi-user load prediction: Train an XGBoost regressor independently for each user i. Its characteristics, in addition to public meteorological and temporal features, also include the user's historical load sequence. Industry-specific electricity consumption characteristics:

[0040] in, The base load forecast value for user i; It contains user-specific historical load information.

[0041] (4) Online residual correction based on sliding window Considering real-time weather changes and short-term fluctuations, a sliding window residual correction mechanism is introduced. Define the past L... Average prediction error over a time period (e.g., L=4, i.e., the past hour):

[0042]

[0043] in, Contribute to practical new energy sources; L represents the actual load of user i; L is the sliding window length (number of time periods).

[0044] The final corrected prediction value is:

[0045]

[0046] in, An adaptive learning rate, ranging from 0.5 to 0.9, is used to control the correction strength. The correction term is updated continuously with actual data, enabling the predicted value to quickly track the true trend. Using the above prediction method, the predicted output of new energy power generation for the next T time periods is output. and the load forecast values ​​for each user (i=1, 2, ..., N).

[0047] Step S3: Multi-objective dynamic multi-user assignment modeling A multi-objective dynamic allocation optimization model is constructed, which includes a two-layer structure of objective function set and constraint condition set.

[0048] (1) Objective function set Objective function F1 (maximizing green electricity consumption): maximize the proportion of self-generated and self-consumed electricity from green electricity direct-connection projects to ensure compliance with policy requirements.

[0049]

[0050] in, Let t be the green electricity power allocated to user i during time period t; t represents the corrected predicted value of renewable energy output during time period t; T represents the total number of time periods in the scheduling cycle (hourly, usually 24); N represents the total number of users.

[0051] Objective function F2 (minimize overall electricity cost): minimize the energy cost of each user and the cost of energy storage over its entire life cycle.

[0052]

[0053] in, Cost per kilowatt-hour for direct green electricity connection; The comprehensive electricity price for purchasing electricity from the grid during time period t (including transmission and distribution price, system operation fee, cross-subsidy, and government fund, in yuan / kWh); The load correction forecast value for user i during time period t; The charging and discharging power of energy stored in time period t (positive value for discharging, negative value for charging). The lifetime degradation cost per unit power charge and discharge of energy storage; The unit time period operation and maintenance cost of energy storage.

[0054] Objective function F3 (carbon emission minimization): Minimize the implicit carbon emissions from purchasing electricity from the grid while ensuring power balance.

[0055]

[0056] in, Let t be the average carbon emission factor of the power grid during the time period t.

[0057] (2) Constraint set Source-load balance constraints:

[0058] in, The power purchased from the grid during time period t; Indicates the energy storage discharge power; This indicates the energy storage charging power.

[0059] Policy-based electricity allocation constraints: Annual self-generated and self-consumed electricity ratio constraint:

[0060] in The minimum self-consumption ratio required by policy (60% in 2026, and no less than 60% for new projects in 2030 and beyond).

[0061] Constraints on the proportion of green electricity in total electricity consumption:

[0062] in This is the minimum green electricity coverage ratio required by policy (currently 30%, with new projects required to have a coverage ratio of no less than 35% from 2030 onwards).

[0063] Internet usage power consumption ratio constraint:

[0064] in The upper limit for the proportion of electricity allowed to be connected to the grid (20%). Power required for energy storage charging. Energy storage SOC constraints:

[0065]

[0066] in, The state of charge (0~1) of the stored energy at the beginning of time period t; , These are the minimum and maximum allowable values ​​for SOC (usually 0.1 and 0.9). The length of a single scheduling period; This refers to the rated capacity of the energy storage. The overall efficiency of energy storage charging and discharging is less than 1. The charging and discharging power of energy stored during time period t (positive value for discharging, negative value for charging).

[0067] User fairness constraints (Gini coefficient constraints):

[0068] in, Gini The Gini coefficient for allocating green electricity shares to each user; s i The percentage of green electricity used by user i; This represents the average percentage of green electricity. G max The maximum permissible Gini coefficient (taken as 0.3~0.4).

[0069]

[0070]

[0071] Personalized constraints for each user: Each user i can set a minimum green energy usage rate. and the highest affordable electricity price :

[0072]

[0073] Step S4: Hybrid optimization solution based on improved non-dominated sorting genetic algorithm (NSGA-III) and multi-agent proximal policy optimization (MAPPO). For the high-dimensional, multi-constraint, multi-objective, time-varying nonlinear optimization problem constructed in step S3, this invention proposes a hierarchical hybrid solution strategy: in the day-ahead scheduling stage, an improved NSGA-III algorithm is used to search for the global Pareto optimal solution set; in the intraday rolling optimization stage, multi-agent near-end policy optimization (MAPPO) is used for real-time dynamic adjustment.

[0074] (1) Day-to-day scheduling - an improved NSGA-III algorithm The scheduling policy is encoded as chromosomes, with each chromosome containing T x N decision variables. and energy storage charging and discharging plans .

[0075] The standard NSGA-III algorithm has the following limitations when solving the aforementioned high-dimensional multi-objective optimization problems: First, random initialization easily leads to the clustering of initial solutions in the decision space, affecting population diversity; second, a fixed set of reference points is difficult to adapt to Pareto fronts of different shapes, resulting in uneven distribution of solutions; and third, the granularity of constraint processing is insufficient, and the sorting of infeasible solutions is relatively coarse, affecting convergence efficiency. To address these limitations, this invention improves upon the algorithm in four aspects: initial population generation, adaptive adjustment of reference points, constraint processing strategies, and local search.

[0076] Improvement 1: Initial population generation based on uniform design. Latin hypercube sampling combined with a uniform design table is used to generate the initial population, ensuring a uniform distribution of initial solutions in the decision space and avoiding the aggregation problem caused by traditional random initialization.

[0077] Improvement 2: Adaptive reference point adjustment mechanism. Traditional NSGA-III relies on a fixed set of reference points. This invention introduces adaptive reference point adjustment: after every 10 iterations, reference points are regenerated using K-means clustering based on the actual distribution of the current Pareto front, ensuring that the reference points always match the shape of the front.

[0078] Improvement 3: Constraint Domination Priority Strategy. For individuals that violate constraints, define a violation degree vector:

[0079] in, For individuals Total constraint violation value; Let be the j-th inequality constraint (after normalization); m is the total number of constraints. In non-dominated sorting, feasible solutions (CV=0) take precedence over infeasible solutions; infeasible solutions are sorted in ascending order of CV value.

[0080] Improvement 4: Local search based on differential evolution. After each generation of NSGA-III evolution, representative individuals on the Pareto front are selected. The selection is based on the uniformity of the target space, that is, the individual that minimizes the variance of the target distance between adjacent individuals on the front is selected. The differential evolution operator is used to perform a local fine search to enhance convergence.

[0081] The four improvements form a collaborative optimization mechanism: Improvement 1 ensures global diversity of the population in the decision space through uniform initialization, laying the foundation for global search; Improvement 2 maintains the uniformity of the Pareto front distribution during evolution through adaptive reference points, maintaining a balance of selection pressure; Improvement 3 enhances the rationality of infeasible solution ranking through a constraint dominance priority strategy, guiding the population to converge towards the feasible region; Improvement 4 accelerates the local convergence accuracy of the Pareto front through local search. These four improvements work synergistically layer by layer, from initial diversity to evolutionary diversity, and from constraint rationality to local convergence accuracy, jointly improving the quality and uniformity of the Pareto front distribution.

[0082] With the above improvements, NSGA-III outputs a set of non-dominated day-ahead scheduling schemes S. Pareto ={π1,π2,....π K Operators can choose the final solution from these options based on their actual preferences (such as prioritizing cost or carbon reduction). .

[0083] (2) Intraday Rolling Optimization – Multi-Agent Proximity Strategy Optimization (MAPPO) Considering the discrepancy between the actual output and load of new energy sources and the day-ahead forecasts, a multi-agent near-field strategy optimization algorithm is used for correction during the intraday phase (rolling every 15 minutes to predict the next 4 hours), such as... Figure 3 .

[0084] Construct four intelligent agents: Agent A1 (Power Agent): Observes the deviation between the actual output of new energy sources and the forecast, as well as real-time meteorological data, and outputs the output adjustment coefficient. .

[0085] Agent A2 (Energy Storage Agent): Observes SOC, electricity price, and predicted supply-demand gap, and outputs adjustments to energy storage charging and discharging power. .

[0086] Agent A3 (allocation agent): Observes the real-time load deviation of each user and their respective set priority weights, and outputs an allocation correction vector. .

[0087] Agent A4 (Balance Agent): Observes total power deviation and outputs grid power purchase adjustment amount. The core mechanism of MAPPO: A centralized training and distributed execution architecture is adopted. During the training phase, the four agents share the global state (including the actual power of all nodes, prediction error, electricity price, carbon factor, etc.), through a centralized value function. Evaluate the value of joint actions. Each agent... k strategy The Proximal Policy Optimization (PPO) update is used, and its global reward function is:

[0088] in, This is the imbalance penalty coefficient; The total power deviation is represented by this reward function. The physical meaning of this reward function is: to measure how close the sum of the joint adjustments made by the four agents is to the total power deviation of the system. The closer the joint adjustments are to the total power deviation, the greater the reward value, thereby guiding the agents to coordinate their actions and ensuring system power balance.

[0089] MAPPO enables intraday rolling optimization to respond quickly to changes in actual operating conditions within milliseconds, correct day-ahead scheduling plans in real time, and output corrected allocation instructions.

[0090] Step S5: Real-time monitoring and closed-loop correction control In actual operation, the data acquisition system continuously monitors power generation and load demand at a high update frequency of 50ms. When the following deviations exceed preset thresholds, a correction control process is initiated: The relative deviation between the actual output of new energy sources and the intraday rolling optimization benchmark value exceeds ±5%; The relative deviation between the actual total load of multiple users and the intraday rolling optimization baseline exceeds ±3%; The energy storage SOC exceeds the preset allowable operating range (0.1~0.9). The real-time electricity price fluctuates significantly (more than ±20%) compared to the forecast. A user experiences an abnormal operating condition such as a sudden increase in emergency load or equipment failure.

[0091] The corrective control employs Model Predictive Control (MPC). Starting from the current operating state, it re-invokes the MAPPO optimizer from step S4 within a short prediction time domain (15-30 minutes) with a 1-minute resolution to generate the next set of corrective control instructions. The optimization objective of MPC is to minimize the cumulative power deviation in the future short time domain.

[0092] in, To predict the number of time-domain steps; The power imbalance at time t+k; These are the control variables (including allocation coefficient, energy storage capacity, and purchased power). Q and R are reference control variables obtained from intraday rolling optimization; Q and R are positive definite weight matrices.

[0093] Through MPC's rolling optimization and feedback correction, a continuous closed-loop optimization of "daily planning - intraday rolling - real-time correction" is achieved.

[0094] Step S6: Blockchain-based storage and traceability of time-sharing green electricity allocation results The allocation results for each time period, electricity consumption data, new energy power generation data, carbon emission reduction, and other information generated in steps S4 and S5 are packaged and uploaded to the consortium blockchain network (using Hyperledger Fabric or FISCO BCOS architecture) to achieve the following: The green electricity allocation records for each user's time period are stored in an immutable manner, meeting the policy compliance requirements for hourly green electricity traceability; To provide credible data evidence for carbon emission accounting and green certificate issuance; During the settlement phase, the green electricity usage ratio and carbon emission factor of each user are automatically summarized to generate a multi-dimensional energy consumption report. Smart contracts automatically execute green electricity fee settlements, reducing human intervention.

[0095] This invention also provides a green electricity direct-connection multi-user intelligent scheduling and load distribution system, which includes the following functional modules: (1) Data acquisition layer: Intelligent metering and monitoring units deployed in new energy power generation units and all user nodes, including: new energy power generation unit electricity meters, intelligent bidirectional metering meters, meteorological data acquisition devices (irradiance meters, anemometers, temperature sensors), which are connected to the central data processing server via a high-speed data bus. ， Outlier detection, missing value imputation, and normalization preprocessing are performed on the collected multi-source heterogeneous data.

[0096] (2) Prediction layer: Deploy the weather classification and XGBoost combined prediction model, and perform the source load prediction and sliding window residual online correction tasks described in step 2.

[0097] (3) Optimize the decision layer: Establish a multi-objective dynamic allocation model, which includes a day-ahead scheduling optimization unit (improved NSGAIII solver) and an intraday rolling optimization unit (MAPPO multi-agent solver), and perform the hybrid optimization calculation described in step four respectively.

[0098] (4) Real-time control layer: includes a real-time data monitoring unit and an MPC correction controller, which performs closed-loop rolling correction by receiving the latest real-time electricity price and supply-demand balance real-time dispatch information from the power trading center.

[0099] (5) User interaction interface layer: Provides visualization display function of system status of each user, supports users to independently set carbon reduction targets and electricity price ceilings and transmit them to the optimization decision layer in real time as personalized constraints for users, and displays green electricity allocation report and carbon emission report to each user.

[0100] (6) Blockchain storage layer: Uploads the optimization decision results, operation data, scheduling instructions and energy efficiency carbon emission quantification reports to the consortium blockchain node. ， The system is equipped with smart contracts that automatically execute green electricity fee settlement and generate hourly green electricity traceability certificates when trigger conditions are met, providing legal and effective technical support for electricity-carbon co-certification.

Claims

1. A method for intelligent scheduling and load allocation in a multi-user green electricity direct connection scenario, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data, including new energy power generation, user load, energy storage system, electricity market and carbon emission data, and perform preprocessing; S2. Perform weather-type clustering on historical meteorological data, and construct new energy output prediction models and load prediction models for each user based on the weather-type results. Obtain source-load prediction values ​​through online correction using sliding window residuals. S3. Based on the source-load prediction values, establish a multi-objective dynamic allocation model with the goals of maximizing green electricity consumption rate, minimizing electricity costs, and minimizing carbon emissions, while satisfying user fairness constraints and user personalization constraints. S4. Use evolutionary computation and multi-agent reinforcement learning for hybrid optimization to obtain green electricity allocation schemes for each user. S5. Monitor the deviation of the green electricity allocation scheme in real time. When preset trigger conditions are met, use model predictive control for closed-loop rolling correction. S6. Upload the corrected allocation results and electricity consumption data to the consortium blockchain, and use smart contracts for green electricity fee settlement and green electricity traceability and evidence storage.

2. The method according to claim 1, characterized in that, In step S1, the new energy power generation data includes photovoltaic output, wind speed and direction, irradiance, ambient temperature and photovoltaic module backsheet temperature; the user load data includes each user's electricity load curve, demand response willingness and adjustable load capacity, and production plan information; the energy storage system data includes energy storage status of charge, charging and discharging power, health status, and temperature; the electricity market data includes real-time electricity price, medium- and long-term contract electricity price, and spot market electricity price forecast; and the carbon emission data includes each user's industry carbon quota, real-time grid carbon emission factor, and carbon trading market price.

3. The method according to claim 1, characterized in that, In step S2, the weather type clustering uses the K-means algorithm to divide historical meteorological data into different weather types, each corresponding to different renewable energy output characteristics and user load patterns. The renewable energy output prediction includes: constructing a feature vector X(t) containing meteorological features, time features, historical statistical features, and weather type unique thermal coding features; training an XGBoost regressor; using X(t) as input, outputting the basic predicted value of renewable energy output; and employing a squared loss function and an early shutdown strategy. The multi-user load prediction includes: independently training an XGBoost regressor for each user; its input features include meteorological features, time features, the user's historical load sequence, and industry electricity consumption characteristic indicators. The industry electricity consumption characteristic indicators include at least the unit output value electricity consumption, interruptible load ratio, and adjustable load ratio.

4. The method according to claim 1, characterized in that, In step S2, the online correction of the sliding window residual specifically involves: calculating the average prediction error over the past L time periods as the correction amount, weighting it with an adaptive learning rate, and then adding it to the source load base prediction value; the correction amount is updated continuously with real-time operating data to dynamically track the actual output and load trends, thereby obtaining hourly corrected prediction values.

5. The method according to claim 1, characterized in that, In step S3, the constraints of the multi-objective allocation optimization model also include source-load balance constraints, policy-based electricity allocation constraints, and energy storage operation constraints; wherein the policy-based electricity allocation constraints include lower limits for self-consumption ratio, lower limits for green electricity coverage ratio, and upper limits for grid-connected electricity ratio; and the energy storage operation constraints include energy storage state-of-charge operation range constraints and energy storage state-of-charge time-series recursive constraints.

6. The method according to claim 1, characterized in that, In step S3, the user fairness constraint uses the Gini coefficient to quantify the fairness of each user's green electricity allocation share. The green electricity allocation share is the proportion of green electricity allocated to each user to the user's total load, and the Gini coefficient does not exceed the preset maximum allowable threshold. The user-personalized constraint includes the user's minimum green electricity utilization rate constraint and the user's maximum affordable electricity price constraint. The constraint parameters can be set by each user and transmitted to the allocation optimization model in real time to participate in the scheduling calculation.

7. The method according to claim 1, characterized in that, In step S3, the green electricity consumption maximization target optimizes the ratio of the total green electricity power allocated to each user in each time period within the scheduling cycle to the total predicted value of new energy output correction, thereby increasing the self-consumption ratio of green electricity direct connection projects. The electricity cost minimization target, when solving for energy costs, uses the sum of the fixed cost per kilowatt-hour of green electricity direct connection and the dynamic grid purchase cost as the user-side energy cost. The dynamic grid purchase cost is determined based on the real-time changing comprehensive grid price and the net purchased power amount determined by the portion of the load forecast exceeding the green electricity allocated to the user. Furthermore, the electricity cost minimization target simultaneously incorporates the dynamic lifetime affected by the charging and discharging power amplitude on the energy storage cost side. The attenuation cost and fixed operation and maintenance cost measured by operating period are used to form a real-time dynamic allocation mechanism for the entire life cycle cost of energy storage; when solving for carbon emissions, the carbon emission minimization target calculates the total implicit carbon emissions generated by the system purchasing electricity from the external grid throughout the entire time period based on the real-time dynamically updated grid average carbon emission factor and the net purchased power, and takes minimizing this total amount as the optimization direction; the green electricity absorption rate maximization target, the electricity cost minimization target and the carbon emission minimization target are used together as parallel optimization targets for multi-objective optimization, which are used to coordinately optimize the green absorption level, economy and low carbon emissions of the system operation in the day-ahead or real-time scheduling of the green electricity direct connection system, under the premise of ensuring power balance constraints.

8. The method according to claim 1, characterized in that, In step S4, the evolutionary computation employs an improved non-dominated sorting genetic algorithm, NSGA-III, to search for the global Pareto optimal solution set for the current day. The improvements include: using Latin hypercube sampling combined with a uniform design table to generate the initial population, ensuring a uniform distribution of the initial solutions in the decision space; after each preset generation, regenerating adaptive reference points using K-means clustering based on the actual distribution of the current Pareto front, matching the reference points with the front shape; introducing a constraint dominance priority strategy, sorting infeasible solutions by constraint violation value, with feasible solutions taking precedence over infeasible solutions; after each generation of evolution, performing a local search on the Pareto front individuals using a differential evolution operator; and NSGA-III outputting a set of current day scheduling schemes as the baseline scheme for subsequent intraday rolling corrections.

9. The method according to claim 1, characterized in that, In step S4, the multi-agent reinforcement learning uses the multi-agent proximal policy optimization (MAPPO) algorithm to perform intraday real-time dynamic correction of the day-ahead scheduling scheme. This includes: constructing four cooperative agents: power agent, energy storage agent, distribution agent, and balancing agent; adopting a centralized training and distributed execution architecture; the observation space and action space of each agent correspond to the state and control variables of the managed objects, respectively; and setting a cooperative reward item in the global reward function. The cooperative reward item is positively correlated with the degree of system power balance under the joint action of all agents.

10. The method according to claim 1 or 9, characterized in that, In step S5, the preset triggering conditions include one of the following five categories: the relative deviation between the actual output of new energy and the intraday rolling optimization benchmark value exceeds the first percentage threshold; the relative deviation between the actual total load of multiple users and the intraday rolling optimization benchmark value exceeds the second percentage threshold; the energy storage charge status exceeds the preset allowable operating range; and the fluctuation of the real-time electricity price compared with the predicted value exceeds the third percentage threshold. If the change in user load within a preset time window exceeds a preset mutation threshold or if the electrical equipment malfunctions, the closed-loop rolling correction adopts a model predictive control method. In each control cycle, with the current operating state as the initial point, within a preset prediction time domain with a preset resolution as the step size, the optimization objective is to minimize the cumulative power deviation within the prediction time domain. The control scheme obtained by intraday rolling optimization is used as the reference benchmark. The multi-agent reinforcement learning is called again to solve for the optimal control sequence. Only the correction control command at the current moment is executed. In the next control cycle, the above process is repeated according to the updated operating state, realizing a continuous closed loop of day-ahead scheduling, intraday rolling, and real-time correction.

11. The method according to claim 1, characterized in that, In step S6, the consortium blockchain adopts the Hyperledger Fabric or FISCO BCOS architecture, and the smart contract includes: recording the green electricity allocation and actual electricity consumption of each user in each time period on the blockchain for evidence storage to meet the hourly green electricity traceability requirements; automatically executing green electricity fee settlement according to the preset green electricity settlement rules; generating hourly resolution green electricity traceability proof for each user, and providing credible data credentials for carbon emission accounting and green certificate issuance.

12. An intelligent scheduling and load distribution system for multi-user green electricity direct connection scenarios, characterized in that, The method for performing any one of claims 1 to 11 comprises: a data acquisition layer for acquiring the multi-source heterogeneous data and performing the preprocessing; a prediction layer for performing weather type clustering, source-load prediction model construction, and online correction of sliding window residuals; an optimization decision layer for establishing the multi-objective dynamic allocation model and performing the hybrid optimization solution; a real-time control layer for performing real-time deviation monitoring and closed-loop rolling correction; a user interface layer for receiving personalized parameters set by the user and transmitting them to the optimization decision layer in real time as personalized constraints for the user, and displaying green electricity allocation results and carbon emission reports to each user; and a blockchain evidence storage layer for performing green electricity fee settlement and green electricity traceability evidence storage, wherein the smart contract is deployed, and automatically performs green electricity fee settlement and generates hourly green electricity traceability evidence when trigger conditions are met.

Images