A source network load storage multi-time scale scheduling method based on dynamic feedback

Abstract

The application discloses a source network load storage multi-time scale scheduling method based on dynamic feedback, relates to the technical field of power system operation control, and comprises the following steps: acquiring system operation state data, non-source load prediction data and a scheduling plan of a last scheduling period, and constructing an intra-day rolling optimization model; taking the scheduling plan of the last period as a hot start initial solution to input the model, carrying out first stage solving, and obtaining a global theoretical lower bound and a current optimal feasible solution; calculating source load prediction deviation vectors of each period, combining related data to obtain a flip urgency score of energy storage integer variables; selecting a reference number of candidate integer variables according to the score, fixing the remaining integer variables, and constructing a mixed programming sub-model; performing second stage solving with a time limit on the mixed programming sub-model, updating the optimal solution and iterating when the solution is better, and stopping until a preset termination condition is met; outputting a scheduling plan of a current scheduling period, calculating an optimal gap, and performing a hierarchical optimization strategy.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a multi-timescale scheduling method for power sources, grids, loads and storage based on dynamic feedback. Background Technology

[0002] With the continuous increase in the penetration rate of new energy sources, the intraday scheduling of the energy source-grid-load-storage system faces the severe challenge of increased forecast uncertainty. Intraday scheduling is often carried out in 15-minute cycles, and each round requires optimization of the equipment operation plan for the next 4 hours, including decision variables such as energy storage charging and discharging status, controllable unit output, and tie-line switching power. To accelerate the solution process, a hot-start strategy is commonly used in engineering practice, that is, the optimal solution obtained in the previous round of optimization is injected as the initial solution for the current round of solver.

[0003] However, when meteorological conditions change abruptly, such as cloud cover causing a sharp drop in photovoltaic output or sudden changes in wind speed causing significant fluctuations in wind power output, structural differences may emerge between the optimal solution of the previous round and the actual optimal solution of the current round. These differences are primarily reflected in the values ​​of integer variables such as the charging and discharging states of energy storage: in the old scheme, energy storage should have been charging at certain times, but in the new scheme, it needs to be discharging. The initial solution provided by hot start will lock the solver within the neighborhood of the old scheme, causing the solver to get trapped in local optima, converge slowly, or even be unable to escape, ultimately outputting a suboptimal scheduling scheme with high costs.

[0004] Existing solutions mainly include two approaches: one is to abandon warm start and use cold start for solving, but this significantly increases the time consumption and makes it difficult to meet the 15-minute real-time requirement; the other is to run multiple sets of solutions with different initial values ​​in parallel online, but this consumes a lot of computational resources and has high engineering deployment costs. How to retain the acceleration advantages of warm start while effectively identifying and escaping the local optimum trap caused by sudden weather changes, and providing a quantitative assessment of solution quality, is a problem that current intraday scheduling technology urgently needs to solve. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-timescale scheduling method for source-grid-load-storage based on dynamic feedback, which solves the problems of existing technologies being prone to getting trapped in local optima and lacking quantitative evaluation of solution quality during hot start-up under meteorological change scenarios.

[0006] To achieve the above objectives, this invention provides a source-grid-load-storage multi-timescale scheduling method based on dynamic feedback, comprising the following steps: Step 100: Obtain system operation status data for the current scheduling cycle, source load prediction data for the future preset time period, and scheduling plan data for the previous scheduling cycle, and construct an intraday rolling optimization model; System operating status data includes energy storage charge status, controllable unit start-up and shutdown status, and tie-line switching power; Step 200: Set the scheduling plan data of the previous scheduling cycle as the initial solution for hot start and input it into the intraday rolling optimization model. Call the solver to perform the first stage of solving the intraday rolling optimization model and generate the global theoretical lower bound and the current optimal feasible solution. Step 300: Based on the source load prediction data and the source load prediction data of the previous scheduling cycle, calculate the prediction deviation vector for each time period, and calculate the flipping urgency score of the energy storage integer variable for each time period based on the current optimal feasible solution, system operating status data and pre-built scenario knowledge base. Step 400: Set the reference number of candidate integer variables, select the reference number of candidate integer variables according to the order of flipping urgency scores, fix the remaining integer variables in the current optimal feasible solution except for the candidate integer variables, and construct a hybrid programming sub-model that only releases the candidate integer variables. Step 500: Call the solver to perform the second stage of solving the hybrid programming sub-model to obtain an updated feasible solution; if the objective function value of the updated feasible solution is better than the objective function value of the current optimal feasible solution, then replace the current optimal feasible solution with the updated feasible solution, and return to execute steps 300 to 500 for iteration until the preset termination condition is met. The second-stage solution has a second time limit; the preset termination conditions include reaching the maximum number of iterations, running out of solution time, or reaching the number of times the feasible solution is in a steady state; when the solver is called to solve the hybrid programming sub-model and an infeasible solution is obtained, infeasibility repair is performed. Step 600: The latest current optimal feasible solution is output as the scheduling plan for the current scheduling cycle. The optimal gap is calculated based on the global theoretical lower bound and the objective function value of the current optimal feasible solution. The hierarchical optimization strategy is executed according to the optimal gap.

[0007] Therefore, the above-mentioned source-grid-load-storage multi-timescale scheduling method based on dynamic feedback in this invention has the following beneficial effects: (1) Through a two-stage hybrid solution architecture, the first stage uses the optimal lower bound priority strategy to quickly tighten the global theoretical lower bound, providing a mathematical basis for subsequent solution quality evaluation; the second stage uses the local branch search guided by the flip urgency score to effectively escape the local optimum trap while retaining the advantage of hot start acceleration, thus balancing solution speed and solution quality. (2) By integrating four-dimensional information such as net load forecasting deviation, electricity price conversion signal, energy storage state of charge pressure and scenario conflict degree, a flip-up urgency score can be constructed, which can accurately identify the integer variables that need to be re-optimized due to sudden weather changes and avoid blind search; (3) The system adaptively maintains the scenario knowledge base through online clustering, eliminating the need for manual pre-labeling of typical scenarios. The system can automatically adapt to seasonal changes and equipment aging as it accumulates over time, resulting in low maintenance costs. (4) The tabu list mechanism avoids search oscillations caused by repeated flipping of the same integer variable, and the infeasibility repair mechanism ensures that subproblems always have solutions, thus enhancing the robustness and engineering reliability of the algorithm. (5) By using the extreme mutation pre-detection step, the guided search is automatically skipped and the conventional solution is directly performed when the prediction deviation is extremely large, thus avoiding the waste of computing resources when the old solution is completely invalid and ensuring the real-time requirements.

[0008] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0009] Figure 1 This is an overall flowchart of a source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to the present invention. Figure 2 This is a schematic diagram of the construction structure of the hybrid planning sub-model according to an embodiment of the present invention; Figure 3 The following is a performance comparison bar chart of the embodiments of the present invention, wherein (a) is a comparison chart of average solution time; (b) is a comparison chart of optimal gap; (c) is a comparison chart of scheduling cost; and (d) is a comparison chart of average iteration rounds. Detailed Implementation

[0010] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0011] Example 1 This embodiment uses a complete dispatch cycle of a microgrid in an industrial park under a sudden weather change scenario as an example to illustrate the specific implementation process of the present invention. The microgrid is configured as follows: a 500kW photovoltaic power station, a 300kW wind turbine, a 500kWh / 150kW battery energy storage system, and a 200kW gas turbine. It is connected to the main grid via a tie line, with a maximum power purchase capacity of 500kW and a maximum power sale capacity of 300kW. The energy storage charge / discharge efficiency is 95%, and the allowable state of charge (SOC) range is 20% to 90%. The time-of-use electricity price is set as follows: 1.2 yuan / kWh during peak hours (08:00-12:00 and 17:00-21:00), 0.8 yuan / kWh during normal hours (12:00-17:00 and 21:00-23:00), and 0.4 yuan / kWh during off-peak hours (23:00-08:00 the next day). The electricity sale price is 80% of the purchase price. The daily rolling optimization is based on a 15-minute cycle, with each optimization covering 16 time periods over the next 4 hours.

[0012] The experiment selected the scheduling process between 14:00 and 15:00. At 14:05 on the same day, a sudden local severe convective weather occurred, and the photovoltaic output dropped sharply from the predicted value of 420kW to the measured value of 180kW within 15 minutes, which is a typical extreme sudden change scenario.

[0013] Step 100: Acquire data and build an intraday rolling optimization model.

[0014] During the 14:15 dispatch cycle, the system acquires the following data: system operation status data, including the current energy storage SOC of 52.3%, the gas turbine being in operation, and the tie-line exchange power of 120kW purchased in the previous cycle; source and load forecast data for the future preset time period (the next 4 hours, 16 15-minute time periods), i.e., the photovoltaic, wind power, and load power curves predicted based on the latest meteorological information; and dispatch plan data for the previous dispatch cycle (14:00 cycle), i.e., a complete dispatch plan for 16 time periods including the energy storage charging and discharging status, charging and discharging power, gas turbine output, tie-line exchange power, and other decision variables for each time period.

[0015] The constructed intraday rolling optimization model is a mixed-integer linear programming (MILP) model. The objective function is to minimize the total operating cost over the next 4 hours, which includes electricity purchase cost, electricity sales revenue (negative cost), gas turbine fuel cost, energy storage charging and discharging losses and depreciation cost, and wind and solar curtailment penalty cost. Constraints include power balance constraints for each time period, energy storage charging and discharging power constraints, dynamic SOC recursive constraints, gas turbine output upper and lower limits and ramping constraints, and tie-line exchange power constraints. The energy storage charging and discharging state variable u(t) for each time period is a 0 / 1 integer variable, where 1 represents charging and 0 represents discharging, to ensure that energy storage does not physically charge and discharge simultaneously. The model has approximately 200 continuous variables, 16 integer variables, and approximately 500 linear constraints.

[0016] Before executing step 200, the extreme mutation prediction step is performed first. First, the total net load deviation between the current period (14:15) and the previous period (14:00) source-load prediction data is calculated. Net load is the load prediction value minus the new energy output prediction value. The sum of the absolute values ​​of the net load prediction deviation rates for 16 time periods is calculated, resulting in a total deviation of 4.8. Second, the moving average (1.2) and moving standard deviation (0.6) of the net load prediction deviation for the same historical time period (the same 14:15 time in the past 30 days) are obtained within the sliding window. The deviation of the current total deviation from the moving average is calculated, i.e., Z-score = (4.8 - 1.2) / 0.6 = 6.0. The preset statistical threshold is 3.0 (3 times the standard deviation criterion). Since 6.0 > 3.0, it is determined to be an extreme mutation scenario. The system will skip steps 200 to 500 and directly call the solver to perform a regular solution on the complete model within the remaining 14-minute time budget, outputting the result as the final scheduling plan. If the deviation does not exceed the threshold, proceed normally to step 200.

[0017] Step 200: First-stage solution, generating the global theoretical lower bound and the current optimal feasible solution. This step is executed if no extreme mutation pre-detection is triggered. The scheduling plan data from the previous scheduling cycle is set as the initial solution for a warm start. The values ​​of each variable are injected into the model's MIP Start via the solver API. The Gurobi solver is called, and the node selection strategy is set to optimal lower bound priority (Focus=1) to prioritize raising the global theoretical lower bound. The first-stage solution time is limited to 40% of the total 15-minute time window, i.e., 6 minutes. After 6 minutes, the solver pauses and outputs two key results: first, the current global theoretical lower bound is 923.5 yuan, meaning that after 6 minutes of rigorous mathematical proof by the solver, the cost of any feasible solution cannot be lower than this value; second, the current optimal feasible solution, with an objective function value of 956.8 yuan, includes a complete 16-period scheduling scheme.

[0018] Step 300: Calculate the Prediction Deviation Vector and the Reversal Urgency Score. Calculate the prediction deviation vector: For the 16 time periods, first calculate the absolute deviation between the current period's net load prediction and the previous period's. For example, in the first time period (14:15-14:30), the current net load prediction is 180kW, while the previous period's was -50kW, resulting in an absolute deviation of 230kW. Divide the absolute deviation by the preset baseline power value of 500kW (maximum power purchased by the tie line), yielding a deviation rate of 0.46. Then, assign a weighted coefficient w(t) = exp(-0.07·t) that decays over time to obtain a weighted deviation value, forming a 16-dimensional prediction deviation vector. More recent time periods have larger weights; the first time period has a weight of 0.93, and the 16th time period has a weight of 0.33.

[0019] Calculate the flip-over urgency score: For the energy storage integer variable u(t) at each time period, calculate its flip-over urgency score, specifically including: First scoring component: Extracts the corresponding weighted deviation value in the prediction deviation vector for this time period. Second scoring component: Determines whether it falls within the preset time range for peak-valley electricity price transition (30 minutes before and after the transition time). 15:00 is the transition period from normal to peak hours. For the period between 14:30 and 15:30, the second scoring component takes the first preset value of 1.0, and for the rest, it takes the second preset value of 0.0.

[0020] The third scoring component calculates the energy storage SOC pressure value. Following the SOC trajectory of the current optimal feasible solution, if the energy storage is in a charging state during that period, the maximum value is taken between the difference between the SOC and the preset upper limit of 90% of the state of charge, and zero; if it is in a discharging state, the maximum value is taken between the difference between the preset upper limit of the state of charge (20%) and the SOC, and zero. This pressure value quantifies the urgency of the energy storage exceeding its limits; a higher value indicates a greater need for a reversal.

[0021] The fourth scoring component: The current source load prediction data is matched with the scene prototypes in the pre-built scene knowledge base using cosine similarity to find the most similar prototype. The typical charging and discharging state of the corresponding time period under this prototype is extracted. If the state of the current optimal feasible solution for this time period is inconsistent with the typical state, the fourth scoring component takes the third preset value of 1.0; otherwise, it takes the fourth preset value of 0.0.

[0022] Finally, the four scoring components were weighted and summed using weights determined through offline grid search (w1=0.5, w2=0.2, w3=0.2, w4=0.1) to obtain the reversal urgency score for the integer variable of energy storage in each time period. In this period, the scores for time periods 1 to 3 were 0.82, 0.76, and 0.71 respectively, forming a clear high-scoring group, while the scores for the remaining time periods were all below 0.4.

[0023] Online adaptive maintenance of the scenario knowledge base: During the matching process described above, the scenario knowledge base is updated synchronously. Source-load prediction data from the current and previous cycles are extracted to construct a scenario feature vector containing the dimensionality-reduced curve features of the net load for the past 2 hours and the next 4 hours, the current SOC, and the electricity price time period label. The Euclidean distance between this vector and existing prototypes in the database is calculated. If the minimum distance is less than a threshold, the vector information is smoothly updated with a learning rate of 0.1 to the center of the nearest prototype and the probability distribution of the typical charging and discharging states of that prototype for each time period. If the minimum distance exceeds the threshold and the total number of prototypes has not reached the upper limit of 20, a new prototype is created. If the database is full, the least recently updated prototype is removed, and a new prototype is created.

[0024] Step 400: Select candidate integer variables and construct a hybrid programming sub-model.

[0025] First, the urgency scores for the 16 time periods are sorted from highest to lowest to form a score sequence. The difference between adjacent scores is calculated, and the largest difference occurs between the 3rd and 4th time periods (0.71-0.38=0.33). The three integer variables preceding the position of the largest difference are identified as the initial candidate variable set.

[0026] Secondly, the historical solution time record table was queried. The maximum number of integer variables that could be safely solved within the second time limit (2 minutes) was 5. Since 3 were within the limit, these 3 variables (corresponding to the three time periods from 14:15 to 15:00) were selected as candidate integer variables.

[0027] Then, a hybrid programming sub-model is constructed that releases only these 3 candidate integer variables: iterate through all 16 integer variables, and for the 3 variables that belong to the candidate set, keep their original 0 / 1 integer constraints; for the remaining 13 integer variables that do not belong to the candidate set, extract their values ​​from the current optimal feasible solution (e.g., charging state 1 in the 4th period), and forcibly set the upper and lower bounds of the variable to this value to achieve complete fixation.

[0028] During the construction process, a tabu list is maintained. This is the first iteration, and the tabu list is empty. If a variable is added to the tabu list in subsequent iterations, then when selecting candidate variables in step 400, the variable in the tabu list must be forcibly excluded from the candidate set.

[0029] Step 500: Second-stage solution and iteration. The Gurobi solver is invoked to perform the second-stage solution on the constructed hybrid programming sub-model. The second time limit is 2 minutes. At this point, the search space of integer variables in the sub-model is drastically reduced from 65,536 combinations to 8 combinations.

[0030] The solver returns an updated feasible solution. The objective function values ​​of the updated feasible solution and the current best feasible solution are compared: if the new solution has a lower cost, the updated feasible solution replaces the current best feasible solution, and the taboo list is cleared (because the solution structure has changed, the old taboos are invalid). The process returns to step 300, where the score is recalculated based on the new solution, and a new round of iteration begins. If no better solution is found, all three candidate integer variables from this round are added to the taboo list, the unimproved count is incremented by 1, and the process returns to step 400 to reselect the candidate set after excluding taboo variables.

[0031] Infeasibility Fix: If the sub-model solution report is infeasible, the infeasibility fix process is initiated. From the candidate variable set, in ascending order of flipping urgency score, two fixed integer variables are released at a time and re-added to the model. The solution is re-solved. If a feasible solution is obtained with a better cost, the current best feasible solution is updated; otherwise, the original solution is maintained and the current batch of candidate variables is added to the taboo list. This process is repeated until a feasible solution is obtained or all fixed variables are released.

[0032] Iteration Termination Criteria: After each iteration, three preset termination criteria are checked. Criterion 1: Solution time exhausted, i.e., less than 1 minute remaining in the total window; Criterion 2: Reaching the steady-state number of feasible solutions, i.e., no better feasible solution is obtained in 3 consecutive iterations; Criterion 3: Exhaustion of candidate integer variables, i.e., all 16 integer variables are in the taboo list. Iteration terminates when any one of these conditions is met.

[0033] In this embodiment, after two rounds of iteration, the iteration terminates because the taboo list excludes the tried variables and the scores of the remaining variables are all low, resulting in no better solution being obtained consecutively, thus satisfying condition two. The objective function value of the current optimal feasible solution is ultimately optimized to 931.2 yuan.

[0034] Step 600: Output the scheduling plan and execute the hierarchical optimization strategy.

[0035] Extract the global theoretical lower bound c = 923.5 elements obtained from the first stage solution. Extract the objective function value of the current optimal feasible solution at the end of the iteration as the global upper bound C = 931.2 elements. Calculate the optimal gap J = |Cc| / C = |931.2 - 923.5| / 931.2 ≈ 0.83%.

[0036] A tiered optimization strategy is implemented: the first gap threshold is preset to 1%, and the second gap threshold is 5%. If the current J = 0.83% ≤ 1%, the solution quality is deemed acceptable, and the current optimal feasible solution is directly output as the scheduling plan for the current scheduling cycle, generating a qualified quality assessment flag. If J is between 1% and 5%, supplementary solutions are required; if J > 5%, a warning containing the optimal gap value is issued to operators during supplementary solutions.

[0037] Final clearing of the taboo list: After the scheduling plan for the current scheduling cycle is output, the taboo list is cleared to prepare for the next 15-minute scheduling cycle.

[0038] Example 2 To verify the technical effect of the present invention, four sets of comparative experiments were set up during the same mutation period of 14:00-15:00: control group 1 adopted a pure hot start strategy; control group 2 adopted a pure cold start strategy; control group 3 adopted a local branching strategy with random selection of candidate variables; the experimental results are shown in Table 1.

[0039] The specific experimental results are shown in Table 1: Table 1: Comparison of Experimental Results

[0040] For results analysis, please refer to [link / reference]. Figure 3 This embodiment compares the actual performance of the present invention through three sets of control experiments. Figure 3 The four subgraphs in the table correspond to the four key evaluation indicators in Table 1. Figure 3 (a) shows the comparison of average solution time. The control group 2 (pure cold start) took as long as 785 seconds, far exceeding the time window, while the present invention took 487 seconds, which is close to the 423 seconds of the control group 1 (pure hot start), and meets the real-time requirements. Figure 3 (b) shows the optimal gap comparison. The control group 1 is trapped in a local optimum with a gap as high as 3.8%, while the gap of this invention is only 0.8%, and the solution quality is significantly better than the pure hot start strategy. Figure 3 (c) in the figure represents a comparison of scheduling costs for 14:15 cycles. The scheduling cost of this invention is 931.2 yuan, which is 4.8% lower than that of control group 1 (978.3 yuan). Figure 3 In the figure, (d) represents the average number of iterations. This invention effectively suppresses search oscillations through the taboo list mechanism, and the number of iterations is controlled at 2.8 rounds.

[0041] As can be seen from the four sub-figures, this invention accurately guides the solution by flipping the urgency score, achieving an 85% hit rate for flipped variables. The taboo list controls the iteration rounds to 2.8 rounds, effectively suppressing oscillations. The total time for the 14:15 cycle is 487 seconds, completed within a 15-minute window; the optimal gap is 0.8%, and the cost is 931.2 yuan, which is 4.8% lower than that of a pure hot start, achieving a good balance between solution speed and solution quality.

[0042] Example 3 Please see Figure 1 This embodiment uses a detailed process within a complete scheduling cycle as an example to illustrate the specific implementation of the present invention.

[0043] Step 100: Acquire data and build an intraday rolling optimization model.

[0044] During the 14:15 dispatch cycle, the system acquires the following data: system operating status data, including the current energy storage SOC of 52.3%, the gas turbine being in operation, and the tie-line switching power being the 120kW of electricity purchased in the previous cycle; source and load forecast data for the future preset time period, i.e., the photovoltaic, wind power, and load power curves predicted for the next 16 time periods based on the latest meteorological information; and dispatch plan data for the previous dispatch cycle, i.e., the complete dispatch plan for 16 time periods obtained by solving the 14:00 cycle, which includes the values ​​of all decision variables for each time period.

[0045] Specifically, the current scheduling cycle refers to the currently executing 15-minute optimization window, which in this embodiment is the optimization cycle started at 14:15. The energy storage state of charge (SOC) in the system operating status data refers to the percentage of remaining energy in the energy storage battery relative to its rated capacity at 14:15, which is 52.3% in this embodiment, with a rated capacity of 500 kWh. The controllable unit start / stop status refers to whether the gas turbine is currently running or shut down; in this embodiment, it is running. The tie-line exchange power refers to the actual power purchased or sold via the tie-line to the main grid in the previous cycle; in this embodiment, it is 120 kW of purchased power. The scheduling plan data for the previous scheduling cycle includes the values ​​of all decision variables, such as the energy storage charge / discharge status u(t) (values ​​of 0 or 1), energy storage charge / discharge power, gas turbine output, tie-line exchange power, and energy storage SOC sequence for each time period.

[0046] Specifically, the constructed intraday rolling optimization model is a mixed-integer linear programming (MILP) model. The objective function is to minimize the total operating cost over the next 4 hours, including electricity purchase cost, electricity sales revenue (as a negative cost), gas turbine fuel cost, energy storage charging and discharging losses and depreciation cost, and wind and solar curtailment penalty cost. The preferred energy storage depreciation cost coefficient is 0.05 yuan / kWh, and the preferred wind and solar curtailment penalty cost coefficient is 0.6 yuan / kWh. Constraints include power balance constraints for each time period, equipment operation constraints, and network security constraints. The power balance constraint requires that the sum of photovoltaic output, wind power output, gas turbine output, energy storage discharge power, and purchased electricity power in each time period equals the load power, energy storage charging power, and electricity sales power. The energy storage equipment operation constraints include upper and lower limits for charge and discharge power, dynamic recursive SOC constraints (both charge and discharge efficiencies are 95%), upper and lower limits for SOC (lower limit 20%, upper limit 90%), and mutual exclusion constraints for charge and discharge implemented through the 0 / 1 integer variable u(t). When u(t) = 1, charging is allowed and the forced discharge power is 0; when u(t) = 0, discharging is allowed and the forced charging power is 0. The gas turbine operation constraints include upper and lower limits for output (the minimum technical output is preferably 40kW) and ramp rate constraints (the maximum ramp rate is preferably 30kW / 15min). The tie-line exchange power constraints include a maximum power purchase of 500kW and a maximum power sale of 300kW. The model contains approximately 200 continuous variables, 16 integer variables, and approximately 500 linear constraints.

[0047] Extreme mutation pre-detection steps.

[0048] Before proceeding to step 100, an extreme mutation pre-detection is performed. First, the total net load deviation between the current cycle and the previous cycle's source load forecast data is calculated. Net load is the load forecast value minus the new energy output forecast value. The sum of the absolute values ​​of the net load forecast deviation rates for 16 time periods is calculated, resulting in a total deviation of 4.8. Second, the moving average (1.2) and moving standard deviation (0.6) of the net load forecast deviation for the same historical time period (at the same time of 14:15 in the past 30 days) are obtained within the sliding window. The degree of deviation of the current total deviation relative to the moving average is calculated, i.e., Z-score = (4.8 - 1.2) / 0.6 = 6.0. The preset statistical threshold is 3.0 (3 times the standard deviation criterion). Since 6.0 > 3.0, it is determined to be an extreme mutation scenario. The system will skip steps 200 to 500 and directly call the solver to perform a regular solution on the complete model within the remaining 14-minute time budget, and output the result as the final scheduling plan. If the deviation does not exceed the threshold, it will proceed normally to step 200.

[0049] Specifically, the total net load deviation refers to a scalar value obtained by summing the net load forecast deviation rates for each time period. The net load forecast deviation rate is calculated as follows: for each time period, first calculate the absolute value of the difference between the current period's net load forecast value and the previous period's net load forecast value, and then divide it by the preset baseline power value of 500kW (maximum power purchased by the tie line). The sliding window refers to the time range formed by looking back a fixed number of days (preferably 30 days in this embodiment) from the current moment. The moving average and moving standard deviation are updated according to a preset period, which is preferably once a day in this embodiment. The degree of deviation is measured by Z-score, and the preset statistical threshold is preferably 3.0. Conventional solution refers to a balanced global search using the solver's default parameter settings. The remaining time budget is approximately 14 minutes remaining after subtracting the time consumed by the extreme mutation pre-detection step (approximately 1 minute) from the 15-minute time window.

[0050] Step 200: The first stage of solving the problem generates the global theoretical lower bound and the current optimal feasible solution.

[0051] This step is executed if no extreme mutation pre-detection is triggered. The scheduling plan data from the previous scheduling cycle is set as the initial solution for a warm start, and the values ​​of each variable are injected into the model's MIP Start via the solver API. The Gurobi solver is called, and the node selection strategy is set to optimal lower bound priority (Focus=1) to prioritize raising the global theoretical lower bound. The first-stage solution time is limited to 40% of the total 15-minute time window, i.e., 6 minutes. After 6 minutes, the solver pauses and outputs two key results: first, the current global theoretical lower bound is 923.5 yuan; second, the current optimal feasible solution has an objective function value of 956.8 yuan.

[0052] Specifically, the initial solution for a hot start refers to assigning the values ​​of each decision variable from the feasible solution obtained after complete optimization in the previous scheduling cycle to the initial search values ​​of the corresponding variables in the current model one by one through the variable initial value setting interface provided by the solver. The optimal lower bound priority node selection strategy means that the solver always prioritizes the node with the lowest theoretical lower bound in the list of nodes to be explored in the branch and bound search tree for branch exploration, in order to improve the global theoretical lower bound as quickly as possible. The first-stage solution time limit is set to 40% of the total time window (i.e., 6 minutes), which is an overall optimization based on the two-stage time allocation. The current global theoretical lower bound (Best Bound) is the value obtained by the solver taking the minimum relaxation lower bound value among all leaf nodes that have not yet been explored, and has strict mathematical significance. The current optimal feasible solution is the integer feasible solution with the lowest objective function value that satisfies all constraints found by the solver during the 6-minute search process, and it is the starting point for the subsequent second-stage guided local search.

[0053] Step 300: Calculate the prediction bias vector and the flip urgency score.

[0054] Calculate the prediction deviation vector: For 16 time periods, first calculate the absolute deviation between the current period and the previous period's net load prediction value. Divide the absolute deviation by the preset baseline power value of 500kW to obtain the deviation rate. Then assign a weighted deviation value w(t)=exp(-0.07*t) that decays over time to obtain the weighted deviation value, forming a 16-dimensional prediction deviation vector.

[0055] Specifically, the formula for the net load forecast is P(t) = Pl(t) - Pv(t) - Pw(t), where P(t) is the net load forecast (kW) for time period t, Pl(t) is the load forecast (kW), Pv(t) is the photovoltaic output forecast (kW), and Pw(t) is the wind power output forecast (kW). The weighting coefficient that decays over time is calculated using an exponential decay function. The decay coefficient λ is determined by requiring that the weight of time period 16 is not less than 30% of that of time period 1. In this embodiment, λ = 0.07 is used.

[0056] Calculate the urgency score for the reversal: For each time period, calculate the urgency score for the energy storage integer variable u(t), specifically including: First score component: extract the corresponding weighted deviation value in the prediction deviation vector for that time period; Second score component: determine whether it is within the preset time range of peak-valley electricity price conversion (30 minutes before and after the conversion time), if yes, take 1.0, otherwise take 0.0; Third score component: calculate the energy storage SOC pressure value, if the energy storage is in a charging state during that time period, take the maximum value of the difference between SOC and the upper limit 90% and zero, if it is in a discharging state, take the maximum value of the difference between the lower limit 20% and SOC and zero; Fourth score component: perform cosine similarity matching between the current source-load prediction data and the scenario prototype in the scenario knowledge base to find the most similar prototype, if the state of that time period in the current solution is inconsistent with the typical state, take 1.0, otherwise take 0.0. Finally, the four scoring components were weighted and summed using weights (w1=0.5, w2=0.2, w3=0.2, w4=0.1) to obtain the reversal urgency score. During this period, the scores for periods 1 to 3 were 0.82, 0.76, and 0.71 respectively, while the scores for the remaining periods were all below 0.4.

[0057] Please refer to the calculation process corresponding to step 300. Figure 2 The system first performs a comprehensive evaluation using four scoring components across four dimensions. These four components are weighted and summed according to preset weights to obtain the urgency score for each time period. Subsequently, the system sorts the scores in descending order from highest to lowest, automatically identifies the drop points by calculating the differences between adjacent scores, and determines a preliminary set of candidate variables. Simultaneously, a time consistency check is performed, comparing the time limits in the historical solution timetable with the required time of the preliminary candidate variable set to ultimately determine the candidate integer variables that need to be released. After locking, the remaining integer variables are fixed, thus constructing a constrained mixed integer programming sub-model, providing a precise search space for the second stage solution in step 500.

[0058] Specifically, the first scoring component has the highest weight w1=0.5 among the four components, reflecting that prediction bias is the primary basis for judging the necessity of integer variable reversal. The preset time range for peak-valley electricity price conversion refers to a 30-minute time window before and after the moment when the electricity price type changes. The first preset value is preferably 1.0, and the second preset value is preferably 0.0. The energy storage SOC pressure value quantifies the urgency of energy storage exceeding its limits. Cosine similarity matching calculates the cosine value of the angle between the net load curve vector of the current source load prediction data and the center vector of each prototype, and takes the prototype with the largest cosine value as the most similar scenario prototype. The specific method for determining weights in offline grid search is as follows: select 100 sets of historical cases as the training set, traverse all possible combinations with a step size of 0.1 between 0.1 and 1.0, and select the weight combination that makes the average recall rate the highest as the default value. In this embodiment, this set of weights achieves a reversal hit rate of about 85% on a typical test set.

[0059] Online adaptive maintenance of the scenario knowledge base: During the matching process described above, the scenario knowledge base is updated synchronously. Source-load prediction data from the current and previous cycles are extracted to construct a 14-dimensional scenario feature vector containing the dimensionality-reduced curve features of the net load for the past 2 hours and the next 4 hours, the current SOC, and the electricity price time period label. The Euclidean distance between this vector and existing prototypes in the database is calculated. If the minimum distance is less than a threshold, the vector information is smoothly updated with a learning rate of 0.1 to the center of the nearest prototype and the probability distribution of the typical charging and discharging states of that prototype for each time period. If the minimum distance exceeds the threshold and the total number of prototypes has not reached the upper limit of 20, a new prototype is created. If the database is full, the least recently updated prototype is removed, and a new prototype is created.

[0060] Specifically, the construction of the scene feature vector adopts a segmented aggregation approximation method, averaging the values ​​of every two adjacent time periods to reduce the 24-dimensional net load data to 12-dimensional curve features. The current energy storage SOC value and electricity price time period label are then added to form a total of 14-dimensional vector. The preset distance threshold is dynamically determined: the initial value is 50% of the average distance between all existing prototypes in the scene knowledge base, and subsequent updates are recalculated at a fixed period (preferably once a day). The preset learning rate is 0.1, and an exponential moving average update method is used: new center = old center × 0.9 + current feature vector × 0.1. The preset upper limit for the total number of scene prototypes is preferably 20. The elimination strategy involves traversing all existing prototypes and selecting the prototype that has not been updated for the longest time since the current time for elimination. If the cumulative matching count is less than 5% of the average matching count of all prototypes, it is directly removed; if the cumulative matching count is high but there has been no recent update, it is marked as "dormant" and retained on disk for offline analysis but does not participate in online matching.

[0061] Step 400: Select candidate integer variables and construct a hybrid programming sub-model.

[0062] First, the urgency scores for the 16 time periods are sorted from highest to lowest. The difference between adjacent scores is calculated, and the largest difference occurs between the 3rd and 4th time periods (0.71-0.38=0.33). The three integer variables preceding the location of the largest difference are identified as the initial candidate variable set. Second, the historical solution time record table is queried. The maximum number of integer variables that can be safely released within the second time limit (2 minutes) is 5. Since 3 are within the limit, these 3 variables are selected as candidate integer variables. Then, a hybrid programming sub-model is constructed that releases only these 3 candidate integer variables: for the 3 variables belonging to the candidate set, their original 0 / 1 integer constraints are maintained; for the remaining 13 integer variables not belonging to the candidate set, their values ​​are extracted from the current optimal feasible solution, and the upper and lower bounds of the variable are forcibly set to these values ​​to achieve complete fixation. During the construction process, a taboo list is maintained.

[0063] Specifically, "drop" refers to the phenomenon where, after the score sequence is sorted in descending order, a sharp drop in score occurs at a certain position, where the difference between adjacent scores is greater than the differences between other adjacent scores; this sharp drop position is the drop. The historical solution time record table is a mapping structure indexed by the number of integer variables released, K, recording the moving average and standard deviation of the actual solution time of the hybrid programming sub-model under different K values. When querying the upper limit of the number, the maximum K value that satisfies "average time + 1 standard deviation ≤ second time limit" is taken. Operations on fixed integer variables modify both the lower and upper bounds of the variable simultaneously. The taboo list is a set of variable-length integer variable numbers valid within the current scheduling period; it is initialized as an empty set and cleared after the scheduling plan for the current scheduling period is output.

[0064] Specifically, the input for time consistency verification consists of two parts: a preliminary candidate variable set and the corresponding number of variables, and a preset second-stage solution time limit. Verification relies on a historical solution time record table maintained by the system. This table, indexed by the number of released integer variables, stores the moving average and moving standard deviation of historical solution times for the mixed programming sub-model under the corresponding variable size. First, based on the historical solution time record table, the maximum number of integer variables satisfying the condition that "the sum of the moving average solution time and one moving standard deviation does not exceed the second time limit" is calculated and determined as the upper limit for the number of integer variables released. If the number of variables in the preliminary candidate variable set is less than or equal to the upper limit for the number of integer variables released, all variables in the preliminary candidate variable set are retained as final candidate integer variables. If the number of variables in the preliminary candidate variable set is greater than the upper limit for the number of integer variables released, variables are sorted from highest to lowest according to their flipping urgency score, and variables with the same number as the upper limit are selected as final candidate integer variables.

[0065] Step 500: Second stage of solution and iteration.

[0066] The Gurobi solver is invoked to perform the second stage of solving the constructed hybrid programming sub-model, with a time limit of 2 minutes. At this point, the search space for integer variables in the sub-model is drastically reduced from 65,536 combinations to 8 combinations. The solver returns an updated feasible solution. The objective function values ​​of the updated feasible solution and the current best feasible solution are compared: if the new solution has a lower cost, the updated feasible solution replaces the current best feasible solution, the tabu list is cleared, and the process returns to step 300 to start a new round of iteration; if no better solution is obtained, all three candidate integer variables from this round are added to the tabu list, the unimproved count is incremented by 1, and the process returns to step 400 to reselect the candidate set.

[0067] Specifically, the complete model has 16 0 / 1 integer variables with 2^16 = 65536 possible combinations, while the sub-model releases only 3 integer variables, resulting in 2^3 = 8 possible combinations. The criterion for determining whether the updated feasible solution's objective function value is better than the current best feasible solution is that the total cost of the new solution is strictly less than the currently saved best cost value. After each better solution is found, the taboo list is immediately cleared because the energy storage SOC trajectory and power allocation in the new solution have undergone global changes, and the solution structure upon which the old taboo list was based no longer exists. The unimproved count refers to the number of consecutive iterations without obtaining a better feasible solution; it is reset to 0 each time a new solution is better than the current best, and incremented by 1 each time no improvement is obtained.

[0068] Infeasibility Fix: If the sub-model solution report is infeasible, the infeasibility fix process is initiated. From the candidate variable set, in ascending order of flipping urgency score, two fixed integer variables are released at a time and re-added to the model. The solution is re-solved. If a feasible solution is obtained with a better cost, the current best feasible solution is updated; otherwise, the original solution is maintained and the current batch of candidate variables is added to the taboo list. This process is repeated until a feasible solution is obtained or all fixed variables are released.

[0069] Specifically, the direct cause of this infeasibility is that, under the new forecast data, the fixed combinations of integer variables create a conflict with the constraints of the remaining continuous variables. The variables are released in order of their urgency of reversal, from lowest to highest score, prioritizing the release of variables with the lowest reversal necessity while preserving the degrees of freedom for high-scoring variables. Releasing two variables at a time strikes a balance between efficiency and reliability.

[0070] Iteration Termination Condition Determination: After each iteration, three preset termination conditions are checked. Condition 1: Solution time is exhausted, i.e., less than 1 minute remains before the end of the total window; Condition 2: The number of times a feasible solution is in a steady state is reached, i.e., no better feasible solution is obtained after 3 consecutive iterations; Condition 3: Candidate integer variables are exhausted, i.e., all 16 integer variables are in the taboo list. The iteration terminates when any one of these conditions is met. In this embodiment, after two rounds of iteration, condition 2 is met, the iteration terminates, and the objective function value of the current optimal feasible solution is optimized to 931.2 yuan.

[0071] Specifically, the time point setting of "less than 1 minute until the end of the total window" in Condition 1 is to reserve approximately 30 to 45 seconds of necessary operation time for step 600. The setting of "no improvement after 3 consecutive iterations" in Condition 2 is based on search efficiency considerations; if there is no improvement after 3 consecutive rounds of trying different combinations of candidate variables, the marginal benefit of continuing the iteration is extremely low. The condition of "exhaustion of candidate integer variables" in Condition 3 is a natural termination behavior after the search space has been fully explored.

[0072] Step 600: Output the scheduling plan and execute the hierarchical optimization strategy.

[0073] The global theoretical lower bound c = 923.5 elements obtained from the first stage solution is extracted, and the objective function value of the current optimal feasible solution at the end of the iteration is extracted as the global upper bound C = 931.2 elements. The optimal gap J = |Cc| / C = |931.2 - 923.5| / 931.2 ≈ 0.83% is calculated. A hierarchical optimization strategy is executed: the first gap threshold is preset to 1%, and the second gap threshold is preset to 5%. Since the current J = 0.83% ≤ 1%, the solution quality is deemed acceptable, and the current optimal feasible solution is directly output as the scheduling plan for the current scheduling cycle, generating a qualified quality assessment flag. The taboo list is cleared after the scheduling plan for the current scheduling cycle is output.

[0074] Specifically, the Minimum Injection Gap (MIP Gap) is an internationally recognized metric that measures the relative gap between the currently found feasible solution and the theoretically possible optimal solution. The formula is J = |Cc| / C. The preferred threshold for the first gap is 1%, and the preferred threshold for the second gap is 5%. Supplementary solving refers to injecting the current optimal feasible solution as the initial solution for a hot start into the solver. The solver then performs a free, variable-fixed solution on the complete model within the remaining time budget (typically 30 to 45 seconds), using the solver's default equilibrium search strategy. The warning indicator is a structured information record containing the MIP Gap value, the objective function value of the current solution, and the trigger time, which is pushed to the human-machine interface of the scheduling personnel.

[0075] This detailed description fully presents the entire process of the present invention within a single scheduling cycle, from data acquisition, model construction, extreme mutation prediction, first-stage global lower bound narrowing, local branch search guided by flip urgency score, taboo list anti-oscillation, infeasibility repair to final quality assessment output. It verifies the technical effectiveness of the method in balancing solution speed and solution quality under meteorological mutation scenarios.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-timescale scheduling method for source-grid-load-storage based on dynamic feedback, characterized in that, Includes the following steps: Step 100: Obtain system operation status data for the current scheduling cycle, source load prediction data for the future preset time period, and scheduling plan data for the previous scheduling cycle, and construct an intraday rolling optimization model; System operating status data includes energy storage charge status, controllable unit start-up and shutdown status, and tie-line switching power; Step 200: Set the scheduling plan data of the previous scheduling cycle as the initial solution for hot start and input it into the intraday rolling optimization model. Call the solver to perform the first stage of solving the intraday rolling optimization model and generate the global theoretical lower bound and the current optimal feasible solution. Step 300: Based on the source load prediction data and the source load prediction data of the previous scheduling cycle, calculate the prediction deviation vector for each time period, and calculate the flipping urgency score of the energy storage integer variable for each time period based on the current optimal feasible solution, system operating status data and pre-built scenario knowledge base. Step 400: Set the reference number of candidate integer variables, select the reference number of candidate integer variables according to the order of flipping urgency scores, fix the remaining integer variables in the current optimal feasible solution except for the candidate integer variables, and construct a hybrid programming sub-model that only releases the candidate integer variables. Step 500: Call the solver to perform the second stage of solving the hybrid programming sub-model to obtain the updated feasible solution; If the objective function value of the updated feasible solution is better than the objective function value of the current optimal feasible solution, then the updated feasible solution replaces the current optimal feasible solution, and the process returns to steps 300 to 500 for iteration until the preset termination condition is met. The second-stage solution has a second time limit; Preset termination conditions include reaching the maximum number of iterations, running out of solution time, or reaching the number of iterations in a feasible solution steady state; When an infeasible solution is obtained by calling the solver to solve the hybrid programming sub-model, infeasibility repair is performed. Step 600: The latest current optimal feasible solution is output as the scheduling plan for the current scheduling cycle. The optimal gap is calculated based on the global theoretical lower bound and the objective function value of the current optimal feasible solution. The hierarchical optimization strategy is executed according to the optimal gap.

2. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 1, characterized in that, The steps for calculating the prediction bias vector for each time period include: Step 301: For each time period within the future preset time period, obtain the net load forecast value of the current scheduling cycle and the net load forecast value of the previous scheduling cycle. The net load forecast value is the load forecast value minus the new energy output forecast value; Step 302: Calculate the absolute deviation of the net load forecast for each time period; The absolute deviation value is the absolute value of the difference between the net load forecast value of the current scheduling period and the net load forecast value of the previous scheduling period. Step 303: Divide the absolute deviation value of each time period by the preset reference power value to obtain the net load prediction deviation rate of each time period; Step 304: Assign a weighting coefficient that decays over time to the net load prediction deviation rate for each time period to obtain the prediction deviation vector, where the weighting coefficient is larger for time periods closer to the current time.

3. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 2, characterized in that, The urgency score for flipping over includes: Step 311: Extract the corresponding values ​​from the prediction deviation vector for each time period as the first scoring component; Step 312: Obtain time-of-use electricity price data and determine whether each time period is within the preset time range of peak-valley electricity price transition. If so, the second scoring component takes the first preset value; otherwise, it takes the second preset value. Step 313: Based on the energy storage charge and discharge state sequence in the current optimal feasible solution and the energy storage state of charge in the system operation state data, calculate the energy storage state of charge pressure value for each time period, and use it as the third scoring component. If the energy storage is in a charging state during the current period, the maximum value between the difference between the energy storage state of charge and the preset upper limit of the state of charge and zero is taken; if the energy storage is in a discharging state during the current period, the maximum value between the difference between the preset upper limit of the state of charge and the energy storage state of charge and zero is taken. Step 314: Match the source load prediction data of the current scheduling cycle with the scenario prototype in the pre-built scenario knowledge base, extract the typical charging and discharging state of energy storage in each time period under the most similar scenario prototype, compare the energy storage charging and discharging state of the corresponding time period in the current optimal feasible solution with the typical charging and discharging state. If they are inconsistent, the fourth score component takes the third preset value; otherwise, the fourth preset value is taken. Step 315: The first, second, third, and fourth scoring components are weighted and summed to obtain the reversal urgency score of the integer variables of the energy storage charging and discharging state for each time period.

4. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 3, characterized in that, The pre-built scenario knowledge base includes: Step 321: Extract source load prediction data for the current scheduling cycle and the previous scheduling cycle, and construct a scenario feature vector. The scenario feature vector includes net load curve features for preset historical time periods and future preset time periods, as well as the current energy storage charge status and the current electricity price time period label. Step 322: Calculate the feature distance between the scene feature vector and the existing scene prototypes in the scene knowledge base. If the minimum feature distance is less than the preset distance threshold, update the scene feature vector to the feature center of the nearest scene prototype with a preset learning rate, and update the probability of occurrence of typical charging and discharging states of energy storage in each time period under the scene prototype. Step 323: If the minimum feature distance is not less than the preset distance threshold and the total number of current scene prototypes has not reached the preset limit, then create a new scene prototype with the scene feature vector as the initial center. Step 324: If the total number of current scene prototypes has reached the preset limit, remove the least recently updated scene prototype from the scene knowledge base and create a new scene prototype with the scene feature vector as the initial center.

5. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 4, characterized in that, The steps for setting a reference number of candidate integer variables include: Step 401: Sort the reversal urgency scores for each time period in descending order to form a score sequence; Step 402: Calculate the difference between adjacent scores in the score sequence, and determine the integer variables corresponding to all scores before the position where the maximum difference occurs as the preliminary candidate variable set; Step 403: Based on the historical solution time record, query the upper limit of the number of integer variables that can be released within the second time limit. If the number of variables in the preliminary candidate variable set is greater than the upper limit, then select the integer variables in the preliminary candidate variable set whose number is equal to the upper limit as candidate integer variables; otherwise, select all integer variables in the preliminary candidate variable set as candidate integer variables.

6. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 5, characterized in that, The steps for constructing the mixed programming sub-model include: Step 501: Traverse all integer variables in the intra-day rolling optimization model, and if the current integer variable belongs to the candidate integer variables, retain its original integer constraint and value range; Step 502: If the current integer variable does not belong to the candidate integer variables, extract the corresponding value of the integer variable from the current optimal feasible solution, and set both the upper bound and the lower bound of the integer variable to the corresponding value to complete the fixation; Step 503: Set solver parameters for the model after variable fixation is completed.

7. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 6, characterized in that, The steps for executing the hierarchical optimization strategy include: Step 601: Extract the global theoretical lower bound obtained at the end of the first-stage solution; Step 602: Extract the objective function value of the current optimal feasible solution at the termination of iteration as the global upper bound; Step 603: Calculate the optimality gap J; J=|(C-c)| / C; Wherein, J is the optimality gap, C is the global upper bound, and c is the global theoretical lower bound; Step 604: Compare the optimality gap with a gap threshold; When J ≤ the first gap threshold, it is determined that the solution quality is qualified, the current optimal feasible solution is directly output as the scheduling plan of the current scheduling cycle, and a qualified quality assessment identifier is generated; When the first gap threshold < J ≤ the second gap threshold, the current optimal feasible solution is used as the warm start initial solution for supplementary solution, and the optimal feasible solution obtained by supplementary solution is output as the scheduling plan of the current scheduling cycle; When J > the second gap threshold, it is determined that the solution quality is questionable, supplementary solution is executed, a warning identifier containing the value of the optimality gap is generated, and the optimal feasible solution obtained by supplementary solution is output as the scheduling plan of the current scheduling cycle.

8. The source-grid-load-storage multi-timescale scheduling method based on dynamic feedback according to claim 7, characterized in that, Before step 100, a mutation pre-inspection step is further included: Step 001: Calculate the total net load deviation between the source-load forecast data of the current scheduling cycle and the source-load forecast data of the previous scheduling cycle; Step 002: Obtain the moving average and moving standard deviation of the net load forecast deviation in the historical same period within the sliding window as statistical distribution data, wherein the moving average and the moving standard deviation are updated according to a preset cycle; Step 003: Calculate the deviation degree of the total net load deviation relative to the moving average, expressed as a multiple of the moving standard deviation; if the deviation degree exceeds a preset statistical threshold, it is determined as an extreme mutation scenario, the solver is directly invoked to perform conventional solution on the intra-day rolling optimization model within the remaining time budget, and the solution result is output as the scheduling plan of the current scheduling cycle.

9. A multi-time-scale scheduling method for source-grid-load-storage based on dynamic feedback as described in claim 8, characterized in that, Constructing the mixed programming sub-model that only releases candidate integer variables in step 400 further includes: Maintaining a tabu list, which is used to record integer variables that have been selected as candidate integer variables in previous iterations of the current scheduling cycle but have not obtained an improvement in the objective function value after the mixed programming sub-model is solved; each time step 400 is re-executed to select candidate integer variables, the integer variables in the tabu list are excluded from the candidate set, until a better feasible solution is obtained in the iteration to clear the tabu list, or the tabu list is cleared after the scheduling plan of the current scheduling cycle is output.

10. A multi-time-scale scheduling method for source-grid-load-storage based on dynamic feedback according to claim 9, characterized in that, The infeasibility repair steps include: Step 511: Release integer variables from the candidate integer variable set in order of flipping urgency score from low to high, releasing a preset number of integer variables each time, and then re-add the released integer variables to the hybrid programming sub-model; Step 512: Recall the solver to solve the updated hybrid programming sub-model. If a feasible solution is obtained and its objective function value is better than the objective function value of the current best feasible solution, then update the current best feasible solution with the feasible solution; otherwise, keep the current best feasible solution unchanged and add all candidate integer variables of the current round to the tabu list. Step 513: Repeat steps 511 and 512 until all fixed integer variables are released; if no feasible solution is obtained after all fixed integer variables are released, terminate the iteration and maintain the current optimal feasible solution unchanged.

Images