Energy storage optimization scheduling method and system under microgrid priority fusion constraint

Abstract

The invention provides an energy storage optimization scheduling method and system under microgrid priority fusion constraint, and the method comprises the steps: S1, carrying out the reading and preprocessing of multi-source data, carrying out the unified time axis of historical photovoltaic output, user load, future weather, electricity price, demand unit price, and completing the missing value interpolation and abnormal value elimination, and forming a time series data set; s2, setting decision variables, namely charging power, discharging power, maximum charging power, maximum discharging power, energy storage equipment capacity, commercial power purchase unit price and demand unit price; s3, the overall income is maximized under the specified priority, and the expenditure is reduced; s4, establishing a priority constraint, an energy storage hardware constraint, an electrical load constraint and other constraint conditions, and avoiding target scheduling conflicts according to the corresponding constraint conditions; s5, carrying out two-stage MINLP rapid solution, and searching a feasible solution in a short time by using a rapid exploration method; and deep optimization is carried out through a fine mode.

Description

Technical Field

[0001] This invention relates to the field of energy storage dispatch, and in particular to an energy storage optimization dispatch method and system under microgrid priority fusion constraints. Background Technology

[0002] Current microgrid energy storage dispatch strategies have the following drawbacks:

[0003] Multiple objectives conflict: Peak shaving, green energy consumption, demand response, and maximum demand modes operate independently, lacking a coordination mechanism. When multiple modes coexist, traditional strategies cannot dynamically coordinate priorities. Disconnect between forecasting and execution: Day-ahead planning relies on historical data and cannot respond in real-time to photovoltaic fluctuations and load changes. Low solution efficiency: Mixed-integer programming models are computationally complex; optimization at 96 time points takes tens of minutes, making it difficult to meet real-time control requirements. Rigid constraints: Fixed charge / discharge period strategies (peak-valley pricing scheduling) easily lead to deep charging and discharging of energy storage, accelerating battery aging. How to optimize microgrid energy storage operations is a crucial technical issue that urgently requires solutions from those skilled in the art. Summary of the Invention

[0004] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method and system for energy storage optimization scheduling under the priority fusion constraint of microgrids.

[0005] To achieve the above-mentioned objectives of the present invention, the present invention provides a method for energy storage optimization scheduling under microgrid priority fusion constraints, comprising the following steps:

[0006] S1, multi-source data reading and time alignment, complete missing value imputation and outlier removal, and construct a standardized time series dataset;

[0007] S2, pvlib calculates theoretical power, LSTM / XGBoost corrects errors, and generates 96-point predictions for photovoltaics and load;

[0008] S3, establish a hard constraint chain based on the priority of peak shaving and valley filling green electricity demand, and use the improved big-M method to linearize the power constraint;

[0009] S4 employs a two-stage MINLP solution strategy to quickly explore and obtain feasible solutions, and uses a fine-tuning mode to deeply optimize the output.

[0010] S5, the optimal charge and discharge curve is sent to EMS, SOC relays across cycles, and real-time comparison realizes the predictive execution closed loop;

[0011] S6, based on prediction error entropy monitoring to trigger rolling re-optimization, extends the cycle during calm periods, and responds instantly during sudden changes;

[0012] S7, Define P ct Pdt E t Decision variables are used to construct an objective function that maximizes the benefits of peak shaving and valley filling for green electricity demand.

[0013] In a preferred embodiment of the above technical solution, step S2 includes:

[0014] First, the theoretical photovoltaic power is calculated using the PVLib photovoltaic power generation calculation library combined with meteorological irradiance. Then, the error is corrected using a Long Short-Term Memory Neural Network (LSTM) / Extreme Gradient Boosting Tree (XGBoost) to obtain the load power P at time t. GFt On the load side, STL decomposition (seasonal trend decomposition) + LightGBM outputs the load power P at time point t. USEt The prediction results also support multi-time point and multi-scale expansion, and 96 points are selected from 192 data points to improve the prediction reliability; and the online rolling prediction mode performs a sliding window prediction of 96 points forward and 192 points backward every 15 minutes.

[0015] In a preferred embodiment of the above technical solution, step S3 includes:

[0016] S3-1 extends the priority from discrete levels to continuous fuzzy instantaneous membership functions μ. k (t)∈[0,1], k∈{peak reduction, valley filling, green electricity, demand}; μ peak (t) represents the instantaneous membership degree of the peak-shaving priority.

[0017] The grid frequency deviation Δf serves as a switching trigger: Hard constraint mode: when Δf > 0.5Hz or Δf < -0.5Hz, μ is forced. peak (t)∈{0,1}, directly assigned by the scheduling instruction; fuzzy constraint mode: when -0.5Hz≤Δf≤0.5Hz, μ peak (t)∈[0,1] is continuously adjustable;

[0018] At confidence level (1-α), ensure that the charging power meets the fuzzy constraint of peak clipping priority; P r It is a probability measure, used to calculate the likelihood of an event occurring.

[0019] Where δ is the power relaxation margin, and its value is δ = 0.1 × P cmax α is the risk tolerance coefficient, with a value of α=0.05; a small amount of charging is allowed to avoid long-term battery idleness;

[0020] When μ peak When (t)=1, the inequality degenerates into P when the peak is hard-cut. ct ≤0, forced charging power is zero;

[0021] When μ peak When (t)=0, without peak clipping, the inequality degenerates into 0≤δ, which always holds, allowing free charging;

[0022] When 0 < μ peak (t)<1, soft peak clipping is achieved through probability constraints, allowing a small amount of charging but requiring the risk of default;

[0023] S3-2, for the bilinear nonlinear constraint that charging power × discharging power = 0, an improved big-M method is introduced for compactification, defining a binary variable u. t ∈{0,1}, the constraint linearizes to:

[0024] P ct ≤P cmax ⋅u t P dt ≤P dmax ⋅(1−u t );u t ∈{0,1},

[0025] Effective inequalities based on minimum operating power:

[0026] P ct ≤P cmin ⋅u t P dt ≤P dmin ⋅(1−u t );

[0027] S3-3, calculates the time-domain safety margin through the scrolling of the scene tree.

[0028] ;

[0029] Where σ SOC The standard deviation of the predicted state of charge of the battery. τ=1,…,t, where τ is the weighted Euclidean distance between the photovoltaic forecast error and the load forecast error. λ is a weighting coefficient, ranging from 0.5 to 1.0, used to balance the dimensional differences between photovoltaic and load errors.

[0030] In a preferred embodiment of the above technical solution, step S4 includes:

[0031] S4-1 employs a two-stage MINLP fast solution, using a fast exploration method to find feasible solutions in a short time; then, it performs deep optimization through a refined mode; a lightweight graph neural network (GNN) is trained as a surrogate model, with the net load curve, electricity price, and priority label as inputs.

[0032] Wherein, nodes represent grid topology nodes (photovoltaic, load, energy storage, grid connection point); edges represent electrical connections, with weights equal to line impedance; the output is an initial feasible solution that satisfies the constraints. The graph neural network loss function includes a constraint violation penalty term L. viol :

[0033] ;in, E represents the predicted energy storage capacity output by the graph neural network (GNN) during offline training. min This represents the minimum energy storage capacity. Using this initial solution as the hot start point of the algorithm, the number of iterations in the first stage is reduced from an average of 15 to less than 5.

[0034] S4-2, perform dynamic adjustment of the trust region radius. If the objective function value does not improve after three consecutive iterations, the trust region radius in the k-th iteration is reduced to 50% of the previous one, i.e., Δ k =0.5×Δ k-1 If the objective function value improves in three consecutive iterations, the radius is increased to twice the size of the previous iteration, but not exceeding the preset upper limit Δ. max , that is, Δ k =min(2×Δ k-1 ,Δ max In other cases, the radius remains unchanged, i.e., Δ k =Δ k-1 Initial value definition: Δ0 = 0.2 × P cmax (kW), Δ max =0.5×P cmax (kW);

[0035] In each iteration, the search range of variables in the mixed-integer programming is limited, |P ct -P ctk−1 |≤Δ k Balance exploration and development, and avoid wasting computing resources in ineffective areas. k Let Δ be the radius of the trust region in the k-th outer approximation iteration. k-1 The preceding radius of the (k-1)th iteration;

[0036] In S4-3, four parallel threads are started simultaneously in the first stage, using different relaxation parameters θ=0.1, 0.3, 0.5, 0.7 to run the external approximation method and build a solution pool. In the second stage, the optimal feasible solution in the pool is selected to start the fine optimization IPOPT-with-Warm-Start based on the interior point method, and Hessian matrix approximation update (LBFGS) is enabled, which compresses the total solution time from hours to 96 time points within 15 minutes.

[0037] In a preferred embodiment of the above technical solution, step S5 further includes:

[0038] S5-1, When there is a communication delay in the energy management system or a sensor failure, dual-model redundancy estimation is initiated:

[0039] Model A: Ampere-hour integration method + temperature compensation;

[0040] Model B: Temporal convolutional network, inputting the voltage, current, and temperature sequences of the most recent 96 time points, outputting the estimated state of charge of the battery;

[0041] Kalman filtering is used to fuse the results of the two models and the confidence weight is dynamically adjusted. When the current ripple exceeds the threshold, the weight of model B is automatically increased to 70% to avoid ampere-hour integral drift.

[0042] In a preferred embodiment of the above technical solution, step S6 further includes:

[0043] S6-1 defines the cumulative error entropy based on a rolling mechanism of prediction error entropy, rather than a fixed period:

[0044] ;

[0045] ;in Let t be the actual photovoltaic power generation. Let be the expected photovoltaic power generation at time t, and ∑error be the total error.

[0046] When the cumulative error entropy H err (t)>H threshold When, i.e., H threshold =0.8 bits, indicating that the prediction error distribution entropy value exceeds 3 times the standard deviation of the stationary state, and rolling re-optimization must be triggered;

[0047] S6-2, based on blockchain, provides evidence storage and auditing for scheduling instructions. Each issued charge / discharge curve is hashed using a secure hash algorithm (SHA-256) and then stored on the blockchain as evidence, with the execution deviation ΔP recorded. exec Smart contract automatic execution deviation assessment: if |ΔP exec If the penalty exceeds 5% for more than 3 consecutive periods, liability will be traced, and a penalty will be automatically deducted from the operator's security deposit. The on-chain mechanism is as follows: Data on-chain mechanism:

[0048] Hardware layer: The energy storage EMS has a built-in Trusted Execution Environment (TEE) that automatically collects the actual power P. act

[0049] Communication layer: Directly pushed to oracle nodes via Modbus-TCP protocol

[0050] Contract layer: Oracle calls to smart contract interfaces (record) execution (curvehash ,P act ,time stamp The curve hash is paired with the measured value for evidence storage to prevent subsequent denial.

[0051] In a preferred embodiment of the above technical solution, step S7 further includes:

[0052] S7-1 defines the objective as maximizing overall revenue under specified priorities. Revenue includes peak shaving revenue from demand response, valley filling revenue from demand response, green electricity consumption revenue, energy storage discharge revenue, and revenue from reducing maximum demand. The overall cost is represented by the summation process:

[0053] Z = z1 + z2 + z3 + z4 + z5;

[0054] Where z1 represents demand response-peak shaving revenue, z2 represents demand response-valley filling revenue, z3 represents green electricity consumption revenue, z4 represents energy storage charging and discharging revenue, and z5 represents revenue from reducing maximum demand.

[0055] z1 is used for calculating the demand response-peak shaving revenue, z1=XY XF1 +XY XF2 +…XY XFt …+XY XF96 ;XY XFt The demand response-peak shaving revenue at time t is calculated as follows:

[0056] When using unit price calculation, XY XFt =XY XFPt *XF PR ;

[0057] z2 Demand Response - Valley Filling Revenue Calculation: z2=XY TG1 +XY TG2 +…XY TGt …+XY TG96 ;XY TGt The demand response minus the valley filling revenue at time t is calculated as follows:

[0058] 1) When using unit price for calculation, XY TGt =XY TGPt *TG PR ;

[0059] 2) To avoid penalties due to insufficient power, XY TGPt =PC max z2 is a constant 0, where XY TGPt Let TG be the valley filling power at time t. PR The unit price for filling valleys in response to demand;

[0060] z3 Green Energy Consumption Benefit Calculation: z3 = LD1 + LD 2+…LD t …+LD 96 ;LD t The revenue from green energy consumption at time t is calculated as follows: When using the unit price for calculation, LD... t =LD pr *P ct ;

[0061] z4 Energy Storage Charging and Discharging Revenue Calculation: z4 = C1 + C 2+ …C t …+C 96 ;

[0062] C t =0.25*DJ t *P ct -0.25*DJ t *P dt ;

[0063] Among them, DJ t This represents the unit price of electricity purchased from the grid at time t.

[0064] z5 reduces the maximum demand benefit calculation: z5=(XLP) max -P USEmax )*XL price ;

[0065] (XLP max =max(XLP1,XLP2,…XLP t …,XLP 96 );

[0066] P USEmax =max(P USE1 ,P USE2 ,…P USEt …,P USE96 );

[0067] Where P USEt XLP represents the estimated power consumption of the user at time t when there is no energy storage charging or discharging. t XL represents the energy demand after the energy storage charge and discharge at time t. price Indicates the unit price of the required quantity;

[0068] The goal is to maximize overall benefits under specified priorities, including peak shaving benefits from demand response, valley filling benefits from demand response, green energy consumption benefits, energy storage discharge revenue, and benefits from reducing maximum demand. To address the difficulty in solver convergence due to the large dimensional differences between z1-z5, preference vector normalization is performed.

[0069] , where λ k Priority weights and It is dynamically determined based on historical data.

[0070] In a preferred embodiment of the above technical solution, step S7 further includes:

[0071] S7-2, Energy Storage Hardware Constraints: Charge / Discharge Power Constraints

[0072] From charging power 0 to maximum charging power P cmax : 0≤P ct ≤P cmax ;

[0073] Discharge power from 0 to maximum discharge power P dmax : 0≤P dt ≤P dmax ;

[0074] Simultaneous charging and discharging: P ct *P dt =0.

[0075] S7-3, Capacity Constraint: The initial capacity is the capacity at the time step before the model runs: E0 = E init ;

[0076] Capacity increases and decreases with charging and discharging:

[0077] E t =E t-1 +0.25*N c *P ct -0.25*P dt / N d ;

[0078] The maximum and minimum capacity are subject to constraints:

[0079] E max *EA min ≤E t =E max *EA max ;

[0080] Among them, E init N represents the initial capacity. c N represents charging efficiency. d E represents the discharge efficiency. t Indicates real-time capacity, E max Indicates credit limit capacity, EA min EA represents the minimum allowable discharge retention capacity percentage. max This indicates the maximum allowed charging capacity percentage; user power consumption is constrained by the estimated user load for discharge: 0 ≤ P dt ≤P USEt Other constraints include limitations at specific times due to specific issues such as typhoons or holidays: P ct=P dt =0; Specifies a specific power charge / discharge, P ct =P ctA P dt =P dtA ;where P ctA and P dtA For a specified fixed value.

[0081] The present invention also discloses a computer system, comprising:

[0082] processor;

[0083] Memory used to store processor-executable instructions;

[0084] The processor is configured to implement the energy storage optimization scheduling method for microgrid priority fusion constraints as described in any one of claims 1 to 9 when executing the executable instructions.

[0085] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0086] 1. By quantitatively decomposing priorities and constraints, problems can be solved precisely and constraint conflicts can be avoided; the objective function quantitatively integrates demand electricity costs, demand response revenue, green electricity consumption revenue, and peak shaving and valley filling revenue to form a complete feasible solution.

[0087] 2. The policy priority can be modified at any time through the front end, and the policy can be switched at any time in the multi-mode superposition scenario, such as the conversion between demand billing and capacity billing, the conversion between photovoltaic grid connection that can be consumed and non-consumed, and the conversion between demand response and no demand response, which can meet the versatility of microgrids under different priorities.

[0088] 3. Consider the integration of multiple scenarios such as peak shaving and valley filling, green electricity consumption, maximum demand, and demand response to improve the overall microgrid revenue while meeting different objectives.

[0089] 4. Compared with manual intervention processes, this solution achieves full automation from data collection to policy issuance, with a total time of ≤3 minutes and reduced response latency.

[0090] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0091] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0092] Figure 1 This is a schematic diagram of the overall operation of the present invention;

[0093] Figure 2This is a schematic diagram illustrating the construction of different priority models in this invention;

[0094] Figure 3 This is a schematic diagram of the process of this invention;

[0095] Figure 4 This is a log diagram showing the data preprocessing results of step A of the present invention;

[0096] Figure 5 This is a flowchart of the online Bayesian weight dynamic update log for step B(a) of the present invention;

[0097] Figure 6 This is a log diagram of the constraint compaction effect in step C(b) of the present invention;

[0098] Figure 7 This is a log flowchart of the entire process of solving the two-stage problem in step D of this invention;

[0099] Figure 8 This is a flowchart of the closed-loop execution of step E and the blockchain evidence storage log of the present invention;

[0100] Figure 9 This is a block diagram of the dual-model SOC soft measurement log for step E(a) of the present invention. Detailed Implementation

[0101] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0102] The data reading process involves extracting historical photovoltaic (PV) output data from a local database or a specified CSV file. This data is typically in time-series format, with each record containing a precise date and timestamp (e.g., YYYY-MM-DD, HH:MM:SS) and the corresponding PV power output value in kW. To meet the needs of intraday optimized scheduling, the data acquisition time resolution is set to 15 minutes, resulting in 96 data points per day. This data not only reflects the randomness and intermittency of PV power generation under weather conditions but also its inherent diurnal periodicity and seasonal variation. Analysis of long-term historical data identifies key performance indicators of the PV system, such as typical output curves, peak power, and capacity factor, providing rich samples for subsequent physical model parameter calibration and statistical learning model training.

[0103] like Figures 1 to 9 As shown, this invention discloses an intelligent optimization scheduling method for energy storage based on the fusion constraints of different priorities in microgrids, including the following:

[0104] S1 performs multi-source data reading and preprocessing, unifies the time axis of historical photovoltaic output, user load and future weather, electricity price and demand unit price, and completes missing value imputation and outlier removal to form a time series dataset, laying the foundation for subsequent forecasting;

[0105] For example, if meteorological data is collected at 15-minute intervals and electricity meter readings are collected at 5-minute intervals, the timestamps of historical photovoltaic, load, and meteorological data are aligned to resolve the discrepancy in data collection frequency; a confidence weighting function based on Cauchy distribution is introduced for each type of data source.

[0106] ;where x i (t) represents the sampled value of data source i at time t, μ i Let γ be the mean of its sliding window, and let γ be an adjustable scaling parameter. During data fusion, γ is used according to w. i (t) Weighted average, automatically reduces the impact weight of abnormal data sources and improves data confidence;

[0107] Multi-source time-series data are constructed into a heterogeneous graph (nodes = data type, edges = time dependencies). A two-layer graph attention network is used to capture the spatiotemporal correlations across data types. Missing values ​​are inferred through weighted aggregation using a neighbor node attention mechanism, accurately capturing the nonlinear coupling relationship between sudden drops in photovoltaic power output and abrupt temperature changes compared to traditional linear imputation. A lightweight physical information neural network is constructed, taking meteorological parameters and historical power output as input and outputting the theoretical power range. When the measured value exceeds three times the standard deviation of the predicted range, it is marked as a candidate anomaly, and then a secondary discrimination is performed using the isolated forest algorithm to avoid misclassifying real rapid fluctuations in cloud cover as anomalies. Through DTW-G time alignment and PINN physical verification, the data anomaly detection rate is >98%, providing a credibility score w_i(t) for step B and supporting online dynamic Bayesian weight allocation.

[0108] S2 generates a scenario using a hybrid prediction model based on a time-series dataset. First, it uses the PVLib library for calculating photovoltaic power generation combined with meteorological irradiance to calculate the theoretical photovoltaic power. Then, it uses a Long Short-Term Memory (LSTM) neural network / Extreme Gradient Boosting Tree (XGBoost) to correct the error, obtaining the photovoltaic power generation P at time t. GFt On the load side, STL decomposition (seasonal trend decomposition) + LightGBM outputs the load power P at time point t. USEt The prediction results also support multi-time point and multi-scale expansion, and 96 points are selected from 192 data points to improve the prediction reliability; and the online rolling prediction mode performs a sliding window prediction of 96 points forward and 192 points backward every 15 minutes.

[0109] a. Correcting photovoltaic predictions: Define the theoretical power P of pvlib. pvlibLong Short-Term Memory Neural Network Output Error Correction Term ΔP LSTM Extreme gradient boosting tree output correction term ΔP XGB ;

[0110] The weights are dynamically assigned using an online Bayesian model:

[0111] P GFt =P pvlib +w LSTM (t)*ΔP LSTM +w XGB (t)*ΔP XGB ;

[0112] P GFt Let w be the photovoltaic power generation at time t, and the confidence weight w of the error correction term output by the long short-term memory neural network. LSTM (t) and the confidence weight w of the extreme gradient boosting tree output correction term XGB (t) Based on the continuous hierarchical probability score of the model's prediction performance over the past 24 hours, the model is adaptively updated through a Dirichlet distribution to achieve intelligent switching between XGBoost for rainy days and LSTM for sunny days.

[0113] b) Input the 96-point prediction results into a Conditional Wasserstein Generative Adversarial Network (cWGAN) to generate 50 extended scenarios with 192 points that conform to the historical joint distribution. During reverse verification, a bidirectional long short-term memory (LSTM) encoder-decoder is used to encode the 192 points, forcing the decoder output to have a DTW (Dynamic Time Warp) dynamic time regularity (DTW) less than a threshold ϵ compared to the original 96-point predictions. KL =0.1, ensuring consistency between short-term and long-term; if the divergence exceeds the limit, the model self-repair mechanism is triggered, and the model is automatically rolled back to the previous stable version to avoid overtraining;

[0114] c. Add a graph embedding layer before the autoregressive integral moving average-artificial neural network model to construct a knowledge graph from user type, holidays, and weather tags. Generate low-dimensional embedding vectors by embedding the knowledge graph into Entity-Embedding+GCN, which serve as auxiliary inputs to the artificial neural network. This enables the model to capture the complex behavior patterns of industrial users actively reducing loads before typhoon warnings, thus improving the robustness of extreme weather prediction.

[0115] S3, the photovoltaic power generation P at time t GFt and the load power P at time t USEtIt serves as the data source for the objective function and constraints; multi-priority fusion constraints are used for modeling, and peak shaving, valley filling, green electricity consumption, and demand management are set as hard constraint chains in order of priority from high to low. During high-priority periods, opposing power is directly blocked, and low-priority revenue items are forced to zero to avoid target conflicts; at the same time, energy storage power, SOC, and user load absorption hardware boundaries are superimposed to form a mixed integer nonlinear programming model, MINLP model, which takes into account both business logic and equipment physics.

[0116] S3-1 extends the priority from discrete levels to continuous fuzzy instantaneous membership functions μ. k (t)∈[0,1], k∈{peak, valley filling, green electricity, demand}; μ peak (t) represents the instantaneous membership degree of the peak-shaving priority.

[0117] The grid frequency deviation Δf serves as a switching trigger: Hard constraint mode: when Δf > 0.5Hz or Δf < -0.5Hz, μ is forced. peak (t)∈{0,1}, directly assigned by the scheduling instruction; fuzzy constraint mode: when -0.5Hz≤Δf≤0.5Hz, μ peak (t)∈[0,1] is continuously adjustable;

[0118] At confidence level (1-α), ensure that the charging power meets the fuzzy constraint of peak clipping priority; P r It is a probability measure, used to calculate the likelihood of an event occurring.

[0119] Where δ is the power relaxation margin, and its value is δ = 0.1 × P cmax α is the risk tolerance coefficient, with a value of α=0.05; a small amount of charging is allowed to avoid long-term battery idleness;

[0120] When μ peak When (t)=1, the inequality degenerates into P when the peak is hard-cut. ct ≤0, forced charging power is zero;

[0121] When μ peak When (t)=0, without peak clipping, the inequality degenerates into 0≤δ, which always holds, allowing free charging;

[0122] When 0 < μ peak (t)<1, soft peak clipping is achieved through probability constraints, allowing a small amount of charging but requiring the risk of default;

[0123] S3-2, for the bilinear nonlinear constraint that charging power × discharging power = 0, an improved big-M method is introduced for compactification, defining a binary variable u. t ∈{0,1}, linearizing the constraints to: Pct P represents the charging power of the stored energy at time t. dt This represents the discharge power at time t.

[0124] P ct ≤P cmax ⋅u t P dt ≤P dmax ⋅(1−u t );u t ∈{0,1}, based on the effective inequality of minimum operating power: P ct ≤P cmin ⋅u t When SOC t−1 ≤95%, P dt ≤P dmin ⋅(1−u t When SOC t−1 ≥5%; of which P cmin and P dmin P is the minimum start-stop power of the energy storage system. cmax and P dmax The maximum start-stop power of the energy storage system is used. Depending on whether different energy storage manufacturers have built-in capacity protection, those without built-in capacity protection are usually taken as SOC*5%. This eliminates invalid solutions where variables approach zero during linear relaxation, significantly reduces the relaxation gap in linear programming, and improves the solution speed.

[0125] S3-3, traditional battery state of charge constraints do not consider the accumulation of prediction errors. By calculating the rolling time-domain safety margin of the scene tree, this constraint ensures that at a confidence level of 90-95%, the battery state of charge will not exceed the limit due to prediction deviation, thus enhancing the reliability of the strategy.

[0126] ;

[0127] E min The purpose of calculating the root mean square of the cumulative error is to predict the average magnitude of the photovoltaic prediction error and the load prediction error, given the minimum energy value corresponding to the lower limit of SOC. Let E be the actual energy storage capacity at time t. t For robustness constraints, it is required that the current SOC must be higher than E. min In addition, an extra safety margin should be reserved.

[0128] Where σ SOC The standard deviation of the predicted state of charge of the battery. τ=1,…,t, where τ is the weighted Euclidean distance between the photovoltaic forecast error and the load forecast error. λ is a weighting coefficient, ranging from 0.5 to 1.0, used to balance the dimensional differences between photovoltaic and load errors; where Let t be the actual photovoltaic power generation at time t. Let t be the expected photovoltaic power generation at time t; The actual power used by the load at time t. Let t be the expected power consumption of the load at time t. The maximum value at time t is taken as the predicted value at time 96.

[0129] S4. Perform a two-stage MINLP fast solution, using a fast exploration method to find a feasible solution in a short time; then perform deep optimization through a fine mode.

[0130] S4-1, train a lightweight graph neural network (GNN) as a surrogate model, with the input being the net load curve, electricity price, and priority label.

[0131] Wherein, nodes refer to grid topology nodes (photovoltaics, loads, energy storage, grid connection points); edges represent electrical connections, with weights equal to line impedance; node characteristics include net load power, electricity price, and μ. k (t) Membership value; Output layer: Projected onto the constrained hyperplane, forced to satisfy P by the ReLU activation function. ct ≤P cmax ⋅u t P dt ≤P dmax ⋅(1−u t Ensure that the output is a feasible solution.

[0132] Output the initial feasible solution that satisfies the constraints. The numbers in parentheses represent the initial charging power, initial discharging power, and initial energy storage capacity, respectively. The graph neural network loss function includes a constraint violation penalty term L. viol :

[0133] ;

[0134] P ct ·P dt For the simultaneity test of charging and discharging, if both are greater than 0, it means the battery is both charging and discharging at the same time, which is physically impossible. In other words, the bidirectional switch of the inverter can only conduct in one direction at a time, and if their product is greater than 0, it violates the constraint. E represents the predicted energy storage capacity output by the graph neural network (GNN) during offline training. min E represents the minimum energy storage capacity. min - For over-discharge indication, if <E min This indicates that the SOC predicted by GNN is lower than the device's safety threshold, causing BMS to trigger a protective shutdown and the scheduling strategy to fail. Using this initial solution as the algorithm's hot start point reduces the number of iterations in the first stage from an average of 15 to less than 5. violThe GNN training applied to step S4 generates high-quality initial solutions through offline training, accelerating convergence.

[0135] S4-2, Perform dynamic adjustment of the trust region radius.

[0136] If the objective function value does not improve after three consecutive iterations, the trust region radius of the k-th iteration is reduced to 50% of the previous one, i.e., Δ k =0.5×Δ k-1 If the objective function value improves in three consecutive iterations, the radius is increased to twice the size of the previous iteration, but not exceeding the preset upper limit Δ. max , that is, Δ k =min(2×Δ k-1 ,Δ max In other cases, the radius remains unchanged, i.e., Δ k =Δ k-1 Initial value definition: Δ0 = 0.2 × P cmax (kW), Δ max =0.5×P cmax (kW);

[0137] In each iteration, the search range of variables in the mixed-integer programming is limited, |P ct -P ctk−1 |≤Δ k Balance exploration and development, and avoid wasting computing resources in ineffective areas. k Let Δ be the radius of the trust region in the k-th outer approximation iteration. k-1 The preceding radius of the (k-1)th iteration;

[0138] In S4-3, four parallel threads are started simultaneously in the first stage, using different relaxation parameters θ=0.1, 0.3, 0.5, 0.7 to run the external approximation method and build a solution pool. In the second stage, the optimal feasible solution in the pool is selected to start the fine optimization IPOPT-with-Warm-Start based on the interior point method, and Hessian matrix approximation update (LBFGS) is enabled, which compresses the total solution time from hours to 96 time points within 15 minutes.

[0139] 10,000 sets of input (net load curve, electricity price, best class label) pairs are generated by running S3 and S4 steps using historical data, and a pre-trained GNN model is deployed online for hot start.

[0140] Optimal charge-discharge curve The data is transmitted to the EMS via Modbus-TCP, and the battery's State of Charge (SOC) is relayed to ensure continuous operation across cycles. The monthly peak demand is automatically updated. At the same time, the measured power is compared in real time to achieve predictive optimization and execute a feedback closed-loop operation, which improves the renewable energy absorption rate and reduces carbon emissions.

[0141] S5-1, When there is a communication delay in the energy management system or a sensor failure, dual-model redundancy estimation is initiated:

[0142] Model A: Ampere-hour integration method + temperature compensation;

[0143] Model B: Temporal convolutional network, inputting the voltage, current, and temperature sequences of the most recent 96 time points, outputting the estimated state of charge of the battery;

[0144] Kalman filtering is used to fuse the results of the two models and the confidence weight is dynamically adjusted. When the current ripple exceeds the threshold, the weight of model B is automatically increased to 70% to avoid ampere-hour integral drift.

[0145] S6-1 defines the cumulative error entropy based on a rolling mechanism of prediction error entropy, rather than a fixed period:

[0146] ;

[0147] ;in Let t be the actual photovoltaic power generation. Let be the expected photovoltaic power generation at time t, and ∑error be the total error.

[0148] When the cumulative error entropy H err (t)>H threshold When, i.e., H threshold =0.8 bits, indicating that the prediction error distribution entropy value exceeds 3 times the standard deviation of the stationary state, and rolling re-optimization must be triggered;

[0149] S6-2, based on blockchain, provides evidence storage and auditing for scheduling instructions. Each issued charge / discharge curve is hashed using a secure hash algorithm (SHA-256) and then stored on the blockchain as evidence, with the execution deviation ΔP recorded. exec Smart contract automatic execution deviation assessment: if |ΔP exec If the percentage of 5% continues to exceed 3 time points, liability will be traced and penalties will be automatically deducted from the operator's deposit. This design ensures the seriousness of dispatching and solves the trust issue between microgrid operators and energy storage owners.

[0150] Its on-chain mechanism is as follows: Data on-chain mechanism:

[0151] Hardware layer: The energy storage EMS has a built-in Trusted Execution Environment (TEE) that automatically collects the actual power P. act ;

[0152] Communication layer: Directly pushed to the oracle node via Modbus-TCP protocol;

[0153] Contract layer: Oracle calls to smart contract interfaces (record)execution (curve hash ,P act ,time stamp The curve hash is paired with the measured value for evidence storage to prevent subsequent denial.

[0154] Define decision variables, let P ct P represents the charging power of the stored energy at time t. dt Let P represent the discharge power at time t. Since the charge and discharge power is sufficient, we set P to... ct and P dt Let P be a real number. cmax and P dmax These represent the maximum charging and discharging power, respectively. E t DJ represents the actual energy storage capacity at time t; t XL represents the unit price of electricity purchased from the grid at time t (in yuan); price This indicates the unit price (in yuan); the following explanation uses the forecast point 96 as an example.

[0155] S7-1 defines the objective as maximizing overall revenue under specified priorities. Revenue includes peak shaving revenue from demand response, valley filling revenue from demand response, green electricity consumption revenue, energy storage discharge revenue, and revenue from reducing maximum demand. The overall cost is represented by the summation process:

[0156] Z = z1 + z2 + z3 + z4 + z5;

[0157] Where z1 represents demand response-peak shaving revenue, z2 represents demand response-valley filling revenue, z3 represents green electricity consumption revenue, z4 represents energy storage charging and discharging revenue, and z5 represents revenue from reducing maximum demand.

[0158] z1 is used for calculating the demand response-peak shaving revenue, z1=XY XF1 +XY XF2 +…XY XFt …+XY XF96 ;XY XFt The demand response-peak shaving revenue at time t is calculated as follows:

[0159] When using unit price calculation, XY XFt =XY XFPt *XF PR ;

[0160] z2 Demand Response - Valley Filling Revenue Calculation: z2=XY TG1 +XY TG2 +…XY TGt …+XY TG96 ;XY TGt The demand response minus the valley filling revenue at time t is calculated as follows:

[0161] 1) When using unit price for calculation, XY TGt =XY TGPt *TG PR ;

[0162] 2) To avoid penalties due to insufficient power, XY TGPt =PC max z2 is a constant 0, where XY TGPt Let TG be the valley filling power at time t. PR The unit price for filling valleys in response to demand;

[0163] z3 Green Energy Consumption Benefit Calculation: z3 = LD1 + LD 2+ …LD t …+LD 96 ;LD t The revenue from green energy consumption at time t is calculated as follows: When using the unit price for calculation, LD... t =LD pr *P ct ;

[0164] z4 Energy Storage Charging and Discharging Revenue Calculation: z4 = C1 + C 2+ …C t …+C 96 ;

[0165] C t =0.25*DJ t *P ct -0.25*DJ t *P dt ;

[0166] Among them, DJ t This represents the unit price of electricity purchased from the grid at time t.

[0167] z5 reduces the maximum demand benefit calculation: z5=(XLP) max -P USEmax )*XL price ;

[0168] (XLP max =max(XLP1,XLP2,…XLP t …,XLP 96 );

[0169] P USEmax =max(P USE1 ,P USE2 ,…P USEt …,P USE96 );

[0170] Where P USEt XLP represents the estimated power consumption of the user at time t when there is no energy storage charging or discharging.t XL represents the energy demand after the energy storage charge and discharge at time t. price Indicates the unit price of the required quantity;

[0171] The goal is to maximize overall benefits under specified priorities, including peak shaving benefits from demand response, valley filling benefits from demand response, green energy consumption benefits, energy storage discharge revenue, and benefits from reducing maximum demand. To address the difficulty in solver convergence due to the large dimensional differences between z1-z5, preference vector normalization is performed.

[0172] , where λ k Priority weights and It is dynamically determined based on historical data.

[0173] S7-2, Energy Storage Hardware Constraints: Charge / Discharge Power Constraints

[0174] From charging power 0 to maximum charging power P cmax : 0≤P ct ≤P cmax ;

[0175] Discharge power from 0 to maximum discharge power P dmax : 0≤P dt ≤P dmax ;

[0176] Simultaneous charging and discharging: P ct *P dt =0.

[0177] S7-3, Capacity Constraint: The initial capacity is the capacity at the time step before the model runs: E0 = E init ;

[0178] Capacity increases and decreases with charging and discharging:

[0179] E t =E t-1 +0.25*N c *P ct -0.25*P dt / N d ;

[0180] The maximum and minimum capacity are constrained:

[0181] E max *EA min ≤E t =E max *EA max ;

[0182] Among them, E init N represents the initial capacity. c N represents charging efficiency.d E represents the discharge efficiency. t Indicates real-time capacity, E max Indicates credit limit capacity, EA min EA represents the minimum allowable discharge retention capacity percentage. max This indicates the maximum allowed charging capacity percentage.

[0183] The user's electricity consumption is constrained by the estimated user load: 0 ≤ P. dt ≤P USEt ;

[0184] Other constraints include limitations at specific times due to specific issues such as typhoons or holidays: P ct =P dt =0;

[0185] Specify a particular power charge / discharge, P ct =P ctA P dt =P dtA ;where P ctA and P dtA For a specified fixed value.

[0186] like Figure 2 As shown in Figure 3, the microgrid energy storage dispatch model undergoes secondary correction and reconstruction. Real-time operational data on the deviation between actual charging and discharging power and predicted values ​​dynamically adjusts the initial dispatch plan. For example, if the prediction error is large, the system automatically strengthens the correction to ensure the adjusted plan better reflects actual operational needs. Multiple corrected schemes are ranked according to indicators such as economic benefit, green energy consumption rate, and battery safety through multi-dimensional screening. For example, schemes with higher returns are prioritized during peak hours, and schemes with higher green energy utilization rates are prioritized when photovoltaic power fluctuations are severe. The final dispatch command is dynamically determined by combining grid conditions (such as frequency fluctuations) and weather conditions (such as photovoltaic power output stability). For example, peak shaving is enforced when the grid frequency is abnormal, and green energy consumption is prioritized when photovoltaic power fluctuations are severe. The execution effect is monitored in real time; if the actual power deviates too much from the dispatch command, the system automatically triggers re-optimization and reconstruction. Simultaneously, dispatch commands and execution results are recorded using blockchain to ensure transparency and traceability of the process.

[0187] It can be used to generate fusion strategies for any future time point, such as 4 time points, 48 ​​time points, 96 time points, and 192 time points. It can also be used to predict 192 points and take 96 points, simplifying the limiting conditions as needed to meet the requirements of timed or real-time delivery.

[0188] Suppose we have an electricity load curve sampled every 15 minutes: 4 o'clock represents the first hour; 48 o'clock represents the 12th hour; 96 o'clock represents the 24th hour; and 192 o'clock represents the 48th hour. If we predict the value at point 96 from point 192, it means using the data from point 192 to deduce or predict the value at point 96. Even if point 96 was originally known, it can be artificially masked, and the subsequent points can be used to predict it backward, verifying whether the model can look back.

[0189] To better illustrate this technical solution, the following example uses a device with a rated capacity of 215 kWh and a rated power of 105 kW for scheduling. The priority of the device is virtual power plant demand response > maximum demand > green electricity consumption > peak shaving and valley filling. The data is sampled every 15 minutes during the specific scheduling process.

[0190] like Figure 3 As shown, this invention preprocesses multi-source data in step S1, generates a scene using a model in step S2, predicts the results, and selects 96 points from 192 data points to improve prediction reliability; and in the online rolling prediction mode, P is used in step S3. GFt and P USEt It serves as the data source for the objective function and constraints, and uses multi-priority fusion constraints for modeling. It combines the constraint activation state interval of S3, and then performs the optimization solution process of S4. Combined with the blockchain audit process of S6, it sets the threshold according to the assessment exemption in S6, and performs the instruction execution process. It acts on the GNN training in step S4, and generates high-quality initial solutions through offline training, thus accelerating convergence.

[0191] Table 1 shows the sampled data. The data was adjusted according to the actual situation in the reasons for reconstruction in Table 1. The notes indicate that the data was replaced when it was clearly inconsistent with the actual situation.

[0192] Serial Number Predicted load power consumption (kWh) Predicted photovoltaic power generation kWh Predicted grid connection power (kWh) Unit price - yuan Capacity-based electricity price - yuan Predicted power consumption (kW) Historical maximum power kW Charge / discharge power (kW) Charge / discharge type capacity kWh Soc restrictions Restricted charge / discharge types Reasons for refactoring 0 68.75 0 68.75 0.290213 35.2 275 624 0 4.3 2 1 66.75 0 66.75 0.290213 35.2 267 624 0 4.3 2 2 77.75 0 77.75 0.290213 35.2 311 624 0 4.3 2 3 70.75 0 70.75 0.290213 35.2 283 624 0 4.3 2 4 71.25 0 71.25 0.290213 35.2 285 624 0 4.3 2 5 69.25 0 69.25 0.290213 35.2 277 624 0 4.3 2 6 64.75 0 64.75 0.290213 35.2 259 624 0 4.3 2 7 59 0 59 0.290213 35.2 236 624 0 4.3 2 8 62.75 0 62.75 0.290213 35.2 251 624 0 4.3 2 9 57.75 0 57.75 0.290213 35.2 231 624 0 4.3 2 10 65.75 0 65.75 0.290213 35.2 263 624 0 4.3 2 11 54.75 0 54.75 0.290213 35.2 219 624 0 4.3 2 12 51.75 0 51.75 0.290213 35.2 207 624 0 4.3 2 13 54.75 0 54.75 0.290213 35.2 219 624 0 4.3 2 14 56.5 0 56.5 0.290213 35.2 226 624 0 4.3 2 15 60.75 0 60.75 0.290213 35.2 243 624 0 4.3 2 16 58.5 0 58.5 0.290213 35.2 234 624 0 4.3 2 17 60.25 0 60.25 0.290213 35.2 241 624 0 4.3 2 18 57.25 0 57.25 0.290213 35.2 229 624 0 4.3 2 19 55.75 0 55.75 0.290213 35.2 223 624 0 4.3 2 20 55 0 55 0.290213 35.2 220 624 0 4.3 2 21 55 0 55 0.290213 35.2 220 624 0 4.3 2 22 55.5 0 55.5 0.290213 35.2 222 624 -105 charge 30.55 14.21 23 54.5 0 54.5 0.290213 35.2 218 624 -105 charge 56.8 26.42 24 54.25 0 54.25 0.290213 35.2 217 624 -105 charge 83.05 38.63 25 53.75 0 53.75 0.290213 35.2 215 624 -105 charge 109.3 50.84 26 54.75 0 54.75 0.290213 35.2 219 624 -105 charge 135.55 63.05 27 55.75 0 55.75 0.290213 35.2 223 624 -105 charge 161.8 75.26 28 60.75 0 60.75 0.290213 35.2 243 624 -105 charge 188.05 87.47 29 49.25 0 49.25 0.290213 35.2 197 624 -105 charge 214.3 99.67 30 55.72 0.22 55.5 0.290213 35.2 222 624 -105 charge 215 100 Avoid not being able to fill 31 74.89 0.89 74 0.290213 35.2 296 624 0 215 100 32 87.62 2.87 84.75 0.656681 35.2 339 624 0 215 100 33 94.75 0.75 94 0.656681 35.2 376 624 0 215 100 34 95.13 6.88 88.25 0.656681 35.2 353 624 0 215 100 35 94.1 9.85 84.25 0.656681 35.2 337 624 0 215 100 36 94.49 11.49 83 0.656681 35.2 332 624 0 215 100 37 91.15 14.15 77 0.656681 35.2 308 624 0 215 100 38 96.29 18.79 77.5 0.656681 35.2 310 624 0 215 100 39 96.93 22.68 74.25 0.656681 35.2 297 624 0 215 100 40 96.87 26.87 70 0.656681 35.2 280 624 0 215 100 41 93.91 25.91 68 0.656681 35.2 272 624 0 215 100 42 93.99 13.74 80.25 0.656681 35.2 321 624 0 215 100 43 93.23 18.73 74.5 0.656681 35.2 298 624 0 215 100 44 91.59 12.34 79.25 1.033041 35.2 317 624 105 put 188.75 87.79 To avoid photovoltaic power being too strong and unable to discharge completely 45 89.26 23.76 65.5 1.033041 35.2 262 624 105 put 162.5 75.58 To avoid photovoltaic power being too strong and unable to discharge completely 46 83.24 26.74 56.5 1.033041 35.2 226 624 105 put 136.25 63.37 To avoid photovoltaic power being too strong and unable to discharge completely 47 78.77 33.52 45.25 1.033041 35.2 181 624 105 put 110 51.16 To avoid photovoltaic power being too strong and unable to discharge completely 48 65.83 38.33 27.5 1.033041 35.2 110 624 105 put 83.75 38.95 To avoid photovoltaic power being too strong and unable to discharge completely 49 67.34 21.34 46 1.033041 35.2 184 624 105 put 57.5 26.74 To avoid photovoltaic power being too strong and unable to discharge completely 50 74.81 16.81 58 1.033041 35.2 232 624 105 put 31.25 14.53 To avoid photovoltaic power being too strong and unable to discharge completely 51 80.66 17.16 63.5 1.033041 35.2 254 624 105 put 5 2.33 To avoid photovoltaic power being too strong and unable to discharge completely. 52 94.58 29.83 64.75 1.033041 35.2 259 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 53 100.39 37.39 63 1.033041 35.2 252 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely. 54 102.46 33.71 68.75 1.033041 35.2 275 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 55 111.86 26.11 85.75 1.033041 35.2 343 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 56 110.72 34.72 76 1.033041 35.2 304 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 57 101.27 39.27 62 1.033041 35.2 248 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 58 97.78 38.53 59.25 1.033041 35.2 237 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 59 99.19 30.19 69 1.033041 35.2 276 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 60 96.91 31.41 65.5 1.033041 35.2 262 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 61 92.55 20.8 71.75 1.033041 35.2 287 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 62 93.23 17.73 75.5 1.033041 35.2 302 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 63 88.54 17.79 70.75 1.033041 35.2 283 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 64 90.79 13.54 77.25 1.033041 35.2 309 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 65 91.5 12.5 79 1.033041 35.2 316 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 66 90.48 8.48 82 1.033041 35.2 328 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 67 90.66 10.16 80.5 1.033041 35.2 322 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 68 87.42 6.42 81 0.656681 35.2 324 624 0 0 0 69 78.88 3.13 75.75 0.656681 35.2 303 624 0 0 0 70 73.25 0 73.25 0.656681 35.2 293 624 0 0 0 71 130 0 130 0.656681 35.2 520 624 -50 charge 12.5 5.81 Replenish, and meet demand. 72 137.5 0 137.5 0.656681 35.2 550 624 -55 charge 26.25 12.21 Replenish, and meet demand. 73 125 0 125 0.656681 35.2 500 624 -105 charge 52.5 24.42 74 125.5 0 125.5 0.656681 35.2 502 624 -105 charge 78.75 36.63 75 68.5 0 68.5 0.656681 35.2 274 624 -105 charge 105 48.84 76 68.5 0 68.5 0.656681 35.2 274 624 -105 charge 131.25 61.05 77 68.75 0 68.75 0.656681 35.2 275 624 -105 charge 157.5 73.26 78 65.25 0 65.25 0.656681 35.2 261 624 -105 charge 183.75 85.47 79 62.75 0 62.75 0.656681 35.2 251 624 -105 charge 210 97.67 80 60.25 0 60.25 1.033041 35.2 241 624 105 put 183.75 85.47 81 56.5 0 56.5 1.033041 35.2 226 624 105 put 157.5 73.26 82 51.25 0 51.25 1.033041 35.2 205 624 105 put 131.25 61.05 83 53.25 0 53.25 1.033041 35.2 213 624 105 put 105 48.84 84 52 0 52 1.033041 35.2 208 624 105 put 78.75 36.63 85 61.25 0 61.25 1.033041 35.2 245 624 105 put 52.5 24.42 86 75 0 75 1.033041 35.2 300 624 105 put 26.25 12.21 87 97 0 97 1.033041 35.2 388 624 105 put 0 0 88 100.75 0 100.75 0.656681 35.2 403 624 0 0 0 89 99.75 0 99.75 0.656681 35.2 399 624 0 0 0 90 96.75 0 96.75 0.656681 35.2 387 624 0 0 0 91 97 0 97 0.656681 35.2 388 624 0 0 0 92 85 0 85 0.656681 35.2 340 624 0 0 0 93 75.5 0 75.5 0.656681 35.2 302 624 0 0 0 94 74.25 0 74.25 0.656681 35.2 297 624 0 0 0 95 70.25 0 70.25 0.656681 35.2 281 624 0 0 0 96 68.75 0 68.75 0.290213 35.2 275 624 0 0 0 97 66.75 0 66.75 0.290213 35.2 267 624 0 0 0 98 77.75 0 77.75 0.290213 35.2 311 624 0 0 0 99 70.75 0 70.75 0.290213 35.2 283 624 0 0 0 100 71.25 0 71.25 0.290213 35.2 285 624 0 0 0 101 69.25 0 69.25 0.290213 35.2 277 624 0 0 0 102 64.75 0 64.75 0.290213 35.2 259 624 0 0 0 103 59 0 59 0.290213 35.2 236 624 0 0 0 104 62.75 0 62.75 0.290213 35.2 251 624 0 0 0 105 57.75 0 57.75 0.290213 35.2 231 624 0 0 0 106 65.75 0 65.75 0.290213 35.2 263 624 0 0 0 107 54.75 0 54.75 0.290213 35.2 219 624 0 0 0 108 51.75 0 51.75 0.290213 35.2 207 624 0 0 0 109 54.75 0 54.75 0.290213 35.2 219 624 0 0 0 110 56.5 0 56.5 0.290213 35.2 226 624 0 0 0 111 60.75 0 60.75 0.290213 35.2 243 624 0 0 0 112 58.5 0 58.5 0.290213 35.2 234 624 0 0 0 113 60.25 0 60.25 0.290213 35.2 241 624 0 0 0 114 57.25 0 57.25 0.290213 35.2 229 624 0 0 0 115 55.75 0 55.75 0.290213 35.2 223 624 0 0 0 116 55 0 55 0.290213 35.2 220 624 0 0 0 117 55 0 55 0.290213 35.2 220 624 -105 charge 26.25 12.21 Avoid not being able to fill 118 55.5 0 55.5 0.290213 35.2 222 624 -105 charge 52.5 24.42 Avoid not being able to fill 119 54.5 0 54.5 0.290213 35.2 218 624 -105 charge 78.75 36.63 120 54.25 0 54.25 0.290213 35.2 217 624 -105 charge 105 48.84 121 53.75 0 53.75 0.290213 35.2 215 624 -105 charge 131.25 61.05 122 54.75 0 54.75 0.290213 35.2 219 624 -105 charge 157.5 73.26 123 55.75 0 55.75 0.290213 35.2 223 624 -105 charge 183.75 85.47 124 60.75 0 60.75 0.290213 35.2 243 624 -105 charge 210 97.67 125 49.25 0 49.25 0.290213 35.2 197 624 -105 charge 215 100 126 55.7 0.2 55.5 0.290213 35.2 222 624 -105 charge 215 100 127 74.9 0.4 74.5 0.290213 35.2 298 624 0 215 100 128 87.68 3.18 84.5 0.656681 35.2 338 624 0 215 100 129 94.86 0.86 94 0.656681 35.2 376 624 0 215 100 130 95.23 2.48 92.75 0.656681 35.2 371 624 0 215 100 131 93.96 6.71 87.25 0.656681 35.2 349 624 0 215 100 132 94.47 11.47 83 0.656681 35.2 332 624 0 215 100 133 91.11 17.86 73.25 0.656681 35.2 293 624 0 215 100 134 96.34 22.84 73.5 0.656681 35.2 294 624 0 215 100 135 96.94 26.94 70 0.656681 35.2 280 624 0 215 100 136 96.83 26.83 70 0.656681 35.2 280 624 0 215 100 137 93.96 29.96 64 0.656681 35.2 256 624 0 215 100 138 93.89 26.64 67.25 0.656681 35.2 269 624 0 215 100 139 93.31 22.06 71.25 0.656681 35.2 285 624 0 215 100 140 91.36 19.36 72 1.033041 35.2 288 624 105 put 188.75 87.79 141 89.23 42.73 46.5 1.033041 35.2 186 624 105 put 162.5 75.58 142 83.37 38.87 44.5 1.033041 35.2 178 624 105 put 136.25 63.37 143 78.72 34.97 43.75 1.033041 35.2 175 624 105 put 110 51.16 144 39.5 43.47 0 1.033041 35.2 0 624 -16 charge 114 53.02 It charges, and it uses green electricity. 145 63.7 37.2 26.5 1.033041 35.2 106 624 105 put 87.75 40.81 To avoid photovoltaic power being too strong and unable to discharge completely 146 74.89 35.89 39 1.033041 35.2 156 624 105 put 61.5 28.6 To avoid photovoltaic power being too strong and unable to discharge completely 147 80.71 35.21 45.5 1.033041 35.2 182 624 105 put 35.25 16.4 To avoid photovoltaic power being too strong and unable to discharge completely 148 94.67 33.67 61 1.033041 35.2 244 624 105 put 9 4.19 To avoid photovoltaic power being too strong and unable to discharge completely 149 100.38 35.13 65.25 1.033041 35.2 261 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 150 102.47 39.97 62.5 1.033041 35.2 250 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 151 111.79 32.79 79 1.033041 35.2 316 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 152 110.75 28 82.75 1.033041 35.2 331 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 153 101.28 23.28 78 1.033041 35.2 312 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 154 97.81 21.81 76 1.033041 35.2 304 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 155 99.18 18.43 80.75 1.033041 35.2 323 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 156 96.83 14.58 82.25 1.033041 35.2 329 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 157 92.58 20.08 72.5 1.033041 35.2 290 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 158 93.23 13.73 79.5 1.033041 35.2 318 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 159 88.35 17.1 71.25 1.033041 35.2 285 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 160 90.86 12.36 78.5 1.033041 35.2 314 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 161 91.4 10.15 81.25 1.033041 35.2 325 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 162 90.34 6.59 83.75 1.033041 35.2 335 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 163 90.73 6.73 84 1.033041 35.2 336 624 105 put 0 0 To avoid photovoltaic power being too strong and unable to discharge completely 164 87.54 3.04 84.5 0.656681 35.2 338 624 0 0 0 165 78.83 1.08 77.75 0.656681 35.2 311 624 0 0 0 166 73.25 0 73.25 0.656681 35.2 293 624 0 0 0 167 78.35 0.1 78.25 0.656681 35.2 313 624 0 0 0 168 81 0 81 0.656681 35.2 324 624 -105 charge 26.25 12.21 169 77.25 0 77.25 0.656681 35.2 309 624 -105 charge 52.5 24.42 170 75.25 0 75.25 0.656681 35.2 301 624 -105 charge 78.75 36.63 171 68.5 0 68.5 0.656681 35.2 274 624 -105 charge 105 48.84 172 68.5 0 68.5 0.656681 35.2 274 624 -105 charge 131.25 61.05 173 68.75 0 68.75 0.656681 35.2 275 624 -105 charge 157.5 73.26 174 65.25 0 65.25 0.656681 35.2 261 624 -105 charge 183.75 85.47 175 62.75 0 62.75 0.656681 35.2 251 624 -105 charge 210 97.67 176 60.25 0 60.25 1.033041 35.2 241 624 105 put 183.75 85.47 177 56.5 0 56.5 1.033041 35.2 226 624 105 put 157.5 73.26 178 51.25 0 51.25 1.033041 35.2 205 624 105 put 131.25 61.05 179 53.25 0 53.25 1.033041 35.2 213 624 105 put 105 48.84 180 52 0 52 1.033041 35.2 208 624 105 put 78.75 36.63 181 61.25 0 61.25 1.033041 35.2 245 624 105 put 52.5 24.42 182 75 0 75 1.033041 35.2 300 624 105 put 26.25 12.21 183 97 0 97 1.033041 35.2 388 624 105 put 0 0 184 100.75 0 100.75 0.656681 35.2 403 624 0 0 0 185 99.75 0 99.75 0.656681 35.2 399 624 0 0 0 186 96.75 0 96.75 0.656681 35.2 387 624 0 0 0 187 97 0 97 0.656681 35.2 388 624 0 0 0 188 85 0 85 0.656681 35.2 340 624 0 0 0 189 75.5 0 75.5 0.656681 35.2 302 624 0 0 0 190 74.25 0 74.25 0.656681 35.2 297 624 0 0 0 191 70.25 0 70.25 0.656681 35.2 281 624 0 0 0

[0193] A 10MW / 20MWh microgrid system in an industrial park; 5MW of photovoltaic power, 20MWh / 10MW of energy storage system, and a maximum load of 8MW, are connected to the grid.

[0194] Table 2: Physical Equipment and Operational Boundary Parameters

[0195] Parameter symbol Parameter name numerical values unit Data source P_cmax Maximum charging power of energy storage 10.0 MW Equipment nameplate P_dmax Maximum discharge power of energy storage 10.0 MW Equipment nameplate P_cmin Minimum charging power for energy storage 0.2 MW Converter start / stop threshold P_dmin Minimum discharge power of energy storage 0.2 MW Converter start / stop threshold E_max Rated capacity of energy storage 20.0 MWh Battery Specifications E_init Initial SOC capacity 10.0 MWh Actual measured values ​​before operation N_c Charging efficiency 0.92 Dimensionless Factory test report N_d Discharge efficiency 0.93 Dimensionless Factory test report EA_min SOC lower limit ratio 0.10 Dimensionless User settings (battery protection) EA_max SOC upper limit ratio 0.95 Dimensionless User settings (battery protection) ΔSOC_safe SOC safety margin 0.05 Dimensionless Running experience value δ Power relaxation margin 1.0 MW =0.1×P_cmax

[0196] Table 3: Hyperparameters of the prediction model

[0197] Model Components Parameter symbol Parameter value Function Description LSTM hidden_units 128 Number of neurons in hidden layer dropout_rate 0.15 Preventing overfitting learning_rate 0.001 Initial learning rate XGBoost n_estimators 300 Number of trees max_depth 6 Maximum tree depth <![CDATA[α XGB (Dirichlet concentration) Dynamic updates Automatically increases to 0.7 or higher in rainy weather. cWGAN n_scenarios 50 Number of generated scenes ϵ_KL threshold 0.1 KL divergence constraint upper limit

[0198] Table 4: Priority and Constraint Parameters

[0199] Priority k μ_k(t) membership degree Triggering conditions Constraint Effective Period peak clipping 0→1 Grid frequency deviation Δf > 0.5Hz Weekdays 10:00-14:00 valley fills the valley 0→1 Grid frequency deviation Δf < -0.5Hz 23:00-05:00 green electricity [0,1] continuous PGFt-PUSEt>0 Photovoltaic peak season demand [0,1] continuous Forecasted demand > 95% of historical peak 24 / 7 monitoring

[0200] Table 5: Economic Parameters

[0201] Parameter symbol name numerical values unit Remark DJ_t Time-of-use electricity pricing 0.3~1.2 Yuan / kWh According to the peak-valley electricity price table <![CDATA[XL price ]]> Quantity per unit price 40.0 Yuan / kW·month Contract electricity price XFPR Peak shaving response unit price 3.5 Yuan / kWh Demand Response Contract TGPR Valley filling response unit price 1.2 Yuan / kWh Demand Response Contract <![CDATA[LD pr ]]> Green electricity consumption subsidy 0.08 Yuan / kWh Renewable energy subsidies

[0202] Table 6: Solver Configuration Parameters

[0203] stage algorithm solver max_iter Convergence tolerance Computing resources Phase 1 OA CBC (MIP) + IPOPT (NLP) 30 0.1% 4 cores 8GB Phase Two IPOPT IPOPT 50 0.01% 8 cores 16GB warm start GNN-Initializer TensorFlow - - 2 cores 4GB

[0204] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for optimal energy storage scheduling under microgrid priority fusion constraints, characterized in that, include: S1, multi-source data reading and time alignment, complete missing value imputation and outlier removal, and construct a standardized time series dataset; S2, pvlib calculates theoretical power, LSTM / XGBoost corrects errors, and generates 96-point predictions for photovoltaics and load; S3, establish a hard constraint chain based on the priority of peak shaving and valley filling green electricity demand, and use the improved big-M method to linearize the power constraint; S4 employs a two-stage MINLP solution strategy to quickly explore and obtain feasible solutions, and uses a fine-tuning mode to deeply optimize the output. S5, the optimal charge and discharge curve is sent to EMS, SOC relays across cycles, and real-time comparison realizes the predictive execution closed loop; S6, based on prediction error entropy monitoring to trigger rolling re-optimization, extends the cycle during calm periods, and responds instantly during sudden changes; S7, Define P ct P dt E t Decision variables are used to construct an objective function that maximizes the benefits of peak shaving and valley filling for green electricity demand.

2. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, S2 includes: First, the theoretical photovoltaic power is calculated using the PVLib photovoltaic power generation calculation library combined with meteorological irradiance. Then, the error is corrected using a Long Short-Term Memory Neural Network (LSTM) / Extreme Gradient Boosting Tree (XGBoost) to obtain the photovoltaic power generation P at time t. GFt On the load side, the power P used by the load at point t is output via STL decomposition + LightGBM. USEt The prediction results also support multi-time point and multi-scale expansion, and 96 points are selected from 192 data points to improve the prediction reliability; and the online rolling prediction mode performs a sliding window prediction of 96 points forward and 192 points backward every 15 minutes.

3. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, S3 includes: S3-1 extends the priority from discrete levels to continuous fuzzy instantaneous membership functions μ. k (t)∈[0,1], k∈{peak reduction, valley filling, green electricity, demand}; μ peak (t) represents the instantaneous membership degree of the peak-shaving priority. The grid frequency deviation Δf serves as a switching trigger: Hard constraint mode: when Δf > 0.5Hz or Δf < -0.5Hz, μ is forced. peak (t)∈{0,1}, directly assigned by the scheduling instruction; fuzzy constraint mode: when -0.5Hz≤Δf≤0.5Hz, μ peak (t)∈[0,1] is continuously adjustable; At confidence level (1-α), ensure that the charging power meets the fuzzy constraint of peak clipping priority; P r It is a probability measure, used to calculate the likelihood of an event occurring. Where δ is the power relaxation margin, and its value is δ = 0.1 × P cmax α is the risk tolerance coefficient, with a value of α=0.05; a small amount of charging is allowed to avoid long-term battery idleness; When μ peak When (t)=1, the inequality degenerates into P when the peak is hard-cut. ct ≤0, forced charging power is zero; When μ peak When (t)=0, without peak clipping, the inequality degenerates into 0≤δ, which always holds, allowing free charging; When 0 < μ peak (t)<1, soft peak clipping is achieved through probability constraints, allowing a small amount of charging but requiring the risk of default; S3-2, for the bilinear nonlinear constraint that charging power × discharging power = 0, an improved big-M method is introduced for compactification, defining a binary variable u. t ∈{0,1}, the constraint linearizes to: P ct ≤P cmax in t ,P dt ≤P dmax ⋅(1−u t );in t ∈{0,1}; S3-3, calculates the time-domain safety margin through the scrolling of the scene tree. ; Where σ SOC The standard deviation of the predicted state of charge of the battery. τ=1,…,t, where τ is the weighted Euclidean distance between the photovoltaic forecast error and the load forecast error. λ is a weighting coefficient, ranging from 0.5 to 1.0, used to balance the dimensional differences between photovoltaic and load errors.

4. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, S4 includes: S4-1 employs a two-stage MINLP fast solution, using a fast exploration method to find feasible solutions in a short time; then, it performs deep optimization through a refined mode; a lightweight graph neural network (GNN) is trained as a surrogate model, with the net load curve, electricity price, and priority label as inputs. Where nodes are power grid topology nodes; edges represent electrical connections, and weights are line impedances; the output is an initial feasible solution that satisfies the constraints. The graph neural network loss function includes a constraint violation penalty term L. viol : ;in, E represents the predicted energy storage capacity output by the graph neural network (GNN) during offline training. min This represents the minimum energy storage capacity. Using this initial solution as the hot start point of the algorithm, the number of iterations in the first stage is reduced from an average of 15 to less than 5. S4-2, perform dynamic adjustment of the trust region radius. If the objective function value does not improve after three consecutive iterations, the trust region radius in the k-th iteration is reduced to 50% of the previous one, i.e., Δ k =0.5×Δ k-1 If the objective function value improves in three consecutive iterations, the radius is increased to twice the size of the previous iteration, but not exceeding the preset upper limit Δ. max , that is, Δ k =min(2×Δ k-1 ,Δ max In other cases, the radius remains unchanged, i.e., Δ k =Δ k-1 Initial value definition: Δ0 = 0.2 × P cmax (kW), Δ max =0.5×P cmax (kW); In each iteration, the search range of variables for the mixed-integer programming is limited, |P ct -P ctk−1 |≤Δ k Balance exploration and development, and avoid wasting computing resources in ineffective areas. k Let Δ be the radius of the trust region in the k-th outer approximation iteration. k-1 The preceding radius of the (k-1)th iteration; In S4-3, four parallel threads are started simultaneously in the first stage, using different relaxation parameters θ=0.1, 0.3, 0.5, 0.7 to run the external approximation method and build a solution pool. In the second stage, the optimal feasible solution in the pool is selected to start the fine optimization IPOPT-with-Warm-Start based on the interior point method, and Hessian matrix approximation update is enabled, which compresses the total solution time from hours to 96 time points within 15 minutes.

5. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, The S5 also includes: S5-1, When there is a communication delay in the energy management system or a sensor failure, dual-model redundancy estimation is initiated: Model A: Ampere-hour integration method + temperature compensation; Model B: Temporal convolutional network, inputting the voltage, current, and temperature sequences of the most recent 96 time points, outputting the estimated state of charge of the battery; Kalman filtering is used to fuse the results of the two models and the confidence weight is dynamically adjusted. When the current ripple exceeds the threshold, the weight of model B is automatically increased to 70% to avoid ampere-hour integral drift.

6. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, S6 further includes: S6-1 defines the cumulative error entropy based on a rolling mechanism of prediction error entropy, rather than a fixed period: ； ;in Let t be the actual photovoltaic power generation. Let be the expected photovoltaic power generation at time t, and ∑error be the total error. When the cumulative error entropy H err (t)>H threshold When, i.e., H threshold =0.8 bits, indicating that the prediction error distribution entropy value exceeds 3 times the standard deviation of the stationary state, and rolling re-optimization must be triggered; S6-2, based on blockchain, provides evidence storage and auditing for scheduling instructions. Each issued charge / discharge curve is hashed using a secure hash algorithm (SHA-256) and then stored on the blockchain as evidence, with the execution deviation ΔP recorded. exec Smart contract automatic execution deviation assessment: if |ΔP exec If the penalty exceeds 5% for more than 3 consecutive periods, liability will be traced, and a penalty will be automatically deducted from the operator's security deposit; the on-chain mechanism is as follows: Hardware layer: The energy storage EMS has a built-in trusted execution environment that automatically collects the actual power P. act ; Communication layer: Directly pushed to the oracle node via Modbus-TCP protocol; Contract layer: Oracle calls to smart contract interfaces (record) execution (curve hash ,P act ,time stamp The curve hash is paired with the measured value for evidence storage to prevent subsequent denial.

7. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, The S7 also includes: S7-1 defines the objective as maximizing overall revenue under specified priorities. Revenue includes peak shaving revenue from demand response, valley filling revenue from demand response, green electricity consumption revenue, energy storage discharge revenue, and revenue from reducing maximum demand. The overall cost is represented by the summation process: Z = z1 + z2 + z3 + z4 + z5; Where z1 represents demand response-peak shaving revenue, z2 represents demand response-valley filling revenue, z3 represents green electricity consumption revenue, z4 represents energy storage charging and discharging revenue, and z5 represents revenue from reducing maximum demand. z1 is used for calculating the demand response-peak shaving revenue, z1=XY XF1 +XY XF2 +…XY XFt …+XY XF96 ;XY XFt The demand response-peak shaving revenue at time t is calculated as follows: When using unit price calculation, XY XFt =XY XFPt *XF PR ; z2 Demand Response - Valley Filling Revenue Calculation: z2=XY TG1 +XY TG2 +…XY TGt …+XY TG96 ;XY TGt The demand response minus the valley filling revenue at time t is calculated as follows: 1) When using unit price for calculation, XY TGt =XY TGPt *TG PR ; 2) To avoid penalties due to insufficient power, XY TGPt =PC max z2 is a constant 0, where XY TGPt Let TG be the valley filling power at time t. PR The unit price for filling valleys in response to demand; z3 Green Energy Consumption Benefit Calculation: z3 = LD1 + LD 2+ …LD t …+LD 96 ;LD t The revenue from green energy consumption at time t is calculated as follows: When using the unit price for calculation, LD... t =LD pr *P ct ; z4 Energy Storage Charging and Discharging Revenue Calculation: z4 = C1 + C 2+ …C t …+C 96 ; C t =0.25*DJ t *P ct -0.25*DJ t *P dt ; Among them, DJ t This represents the unit price of electricity purchased from the grid at time t; z5 reduces the maximum demand benefit calculation: z5=(XLP) max -P USEmax )*XL price ; (XLP max =max(XLP1,XLP2,…XLP t …,XLP 96 ); P USEmax =max(P USE1 ,P USE2 ,…P USEt …,P USE96 ); Where P USEt XLP represents the estimated power consumption of the user at time t when there is no energy storage charging or discharging. t XL represents the energy demand after the energy storage charge and discharge at time t. price Indicates the unit price of the required quantity; The goal is to maximize overall benefits under specified priorities, including benefits from demand response peak shaving, demand response valley filling, green electricity consumption, energy storage discharge, and reduction of maximum demand. To address the difficulty in solver convergence due to the large dimensional differences between z1-z5, preference vector normalization is implemented. , where λ k Priority weights and It is dynamically determined based on historical data.

8. The energy storage optimization scheduling method under microgrid priority fusion constraints according to claim 1, characterized in that, The S7 also includes: S7-2, Energy Storage Hardware Constraints: Charge / Discharge Power Constraints From charging power 0 to maximum charging power P cmax : 0≤P ct ≤P cmax ; Discharge power from 0 to maximum discharge power P dmax : 0≤P dt ≤P dmax ; Simultaneous charging and discharging: P ct *P dt =0; S7-3, Capacity Constraint: The initial capacity is the capacity at the time step before the model runs: E0 = E init ; Capacity increases and decreases with charging and discharging: E t =E t-1 +0.25*N c *P ct -0.25*P dt / N d ; The maximum and minimum capacity are subject to constraints: Yes max *YES min ≤Yes t =Yes max *YES max ; Among them, E init N represents the initial capacity. c N represents charging efficiency. d E represents the discharge efficiency. t Indicates real-time capacity, E max Indicates credit limit capacity, EA min EA represents the minimum allowable discharge retention capacity percentage. max This indicates the maximum allowed charging capacity percentage; user power consumption is constrained by the estimated user load for discharge: 0 ≤ P dt ≤P USEt ; The specified time point is restricted: P ct =P dt =0; Specifies a specific power charge / discharge, P ct =P ctA P dt =P dtA ;where P ctA and P dtA For a specified fixed value.

9. A computer system, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the energy storage optimization scheduling method for microgrid priority fusion constraints as described in any one of claims 1 to 8 when executing the executable instructions.

Images