A power supply and demand balance optimization system based on source network load cooperation

Abstract

This invention relates to the field of energy management technology, specifically disclosing a power supply and demand balance optimization system based on source-grid-load coordination. The system includes a data acquisition and prediction module, which acquires data from the distribution network and the external environment, and outputs prediction curves and confidence intervals; a power flow feasible region construction module, which generates a power flow feasible region based on the prediction curves and confidence intervals, and outputs a time-based constraint set; and a multi-time-scale rolling optimization module, which, within the power flow feasible region, simultaneously solves the time-based constraint set and uses a feasible region projection time-series scheduling algorithm for online rolling optimization; when the voltage proximity index is detected to exceed a preset threshold, a short-window re-optimization is initiated to optimize the power supply and demand balance. This invention satisfies network security and power supply and demand balance in distribution network scenarios with high penetration of distributed energy and coexisting uncertainties.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, and more specifically, to a power supply and demand balance optimization system based on source-grid-load coordination. Background Technology

[0002] With the increasing integration of distributed power sources (such as photovoltaic, wind power, and gas-fired distributed generators), electrochemical energy storage, and interruptible loads into the distribution network, multi-entity coordination among power sources, grid, load, and storage has become the new normal for balancing power supply and demand. Compared to the traditional dispatching model based on unidirectional power flow and deterministic loads, the current distribution network exhibits characteristics such as strong output volatility, significant spatiotemporal imbalance, and tightly coupled operational constraints: node voltages are significantly affected by local injections and reactive power regulation; thermal stability constraints of lines and main transformers are more easily triggered during local peak periods; and the state of charge (SOC) of energy storage is inherently coupled with cross-period power balance, reserve coverage, and ramp-up capabilities. These factors collectively increase the difficulty of supply and demand coordination and the risks to safe and economical operation at the distribution network level.

[0003] On the one hand, data sources from the distribution side and the external environment are fragmented, inconsistent in granularity, and vary in quality. In engineering, it is often necessary to collect multi-source data from distribution transformers, feeders, distributed power inverters and energy storage converters, distribution automation terminals, and smart meters through the interface between the station control layer and the field layer, according to protocols such as IEC 61850, IEC 60870-5-104, and OPC UA. At the same time, time-of-use pricing and meteorological elements (irradiance, temperature, wind speed and direction, etc.) are also connected. Due to the objective existence of inconsistent measurement frequencies (seconds / minutes / 15 minutes / hours), clock drift, missing measurements and outliers, and noise pollution, it is necessary to perform clock synchronization, missing data interpolation, anomaly removal, noise reduction and resampling, and complete node / feeder mapping and cross-scale alignment. Otherwise, the prediction accuracy and the feasibility of subsequent optimization will be directly affected.

[0004] On the other hand, current forecasting practices in engineering still primarily rely on point-based forecasts. While exponential smoothing, ARIMA family of algorithms, or regression methods incorporating dummy variables such as weather, holidays, and time periods can improve results, the characterization of forecast uncertainty often remains at the level of empirical margins or static safety factors, lacking a description of confidence intervals based on historical residual distributions or quantile modeling. This lack of quantification of uncertainty boundaries can lead to two extremes during the optimization phase: first, excessive conservatism to avoid exceeding limits, resulting in unnecessary power curtailment and increased costs; second, planning based on single-point values ​​while ignoring tail risks, leading to voltage and current limits being exceeded and requiring emergency response during actual operation.

[0005] Existing literature (Mathematical Model for Day-ahead Optimal Scheduling of Microgrids, 2017) focuses on day-ahead (single time scale) optimal scheduling of microgrids, centering on the coordinated combination of wind power, photovoltaic, battery storage, and grid power purchase and sale. It first establishes three types of models based on supplementary data: time-of-use calculation model, single-objective integer programming, and single-objective nonlinear optimization. Binary enumeration and MATLAB / Lingo are used for solving these models, comparing the power supply composition, total daily power supply cost, and average electricity purchase price under different scenarios. However, this model primarily relies on static optimization at a single time scale, solving within a fixed window and issuing plans all at once. It lacks feedback on rolling updates to forecasts and state evolution, making it difficult to absorb new information and correct deviations in a timely manner. This leads to frequent occurrences of cross-window power imbalances and execution violations (ramp-up, minimum start-stop).

[0006] When the operating point approaches the constraint boundary, existing solutions mostly rely on manual intervention or rule-triggered unit-level adjustments. They lack a re-optimization closed loop that simultaneously considers ramping, minimum start-stop, backup coverage, and SOC boundary within a short window (such as 5–10 minutes). The correction is lagging, the end-to-end time delay is large, and there is a lack of a fast correction mechanism, resulting in an imbalance between power supply and demand.

[0007] In summary, it can be seen that in the context of distribution networks with high penetration of distributed energy and uncertainty, existing scheduling methods based on single time scale and post-verification are difficult to simultaneously meet the comprehensive requirements of network security and power supply and demand balance. Summary of the Invention

[0008] To overcome the aforementioned deficiencies in the prior art, this invention provides a power supply and demand balance optimization system based on source-grid-load coordination. This system generates a feasible power flow domain based on predicted curves and confidence intervals, unifies it onto a rolling window time axis, and employs a feasible domain projection time-series scheduling algorithm. Through online rolling solutions, it optimizes the power supply and demand balance. When the feasible domain projection time-series scheduling algorithm detects online that the voltage proximity index exceeds a preset threshold, it initiates a short-window re-optimization process to ensure that the operating point remains stably within the feasible power flow domain and maintains a coordinated balance between power and electricity, thereby solving the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A power supply and demand balance optimization system based on source-grid-load coordination includes: The data acquisition and prediction module acquires data from the power distribution network and the external environment, and outputs prediction curves and confidence intervals. The power flow feasible region construction module generates the power flow feasible region based on the predicted curve and confidence interval, and outputs a time-based constraint set; The multi-timescale rolling optimization module, within the feasible domain of power flow, combines time-based constraint sets and solves the problem online using a feasible domain projection time-series scheduling algorithm. If the voltage proximity index exceeds a preset threshold, a short window is initiated for further optimization to maintain the balance between power supply and demand.

[0010] As a further aspect of this invention, the data acquisition and prediction module acquires data from the distribution network and the external environment, and outputs prediction curves and confidence intervals. Specifically, the data acquisition and prediction module acquires and organizes data such as distribution transformer / feeder measurements, distributed power generation, load, price, and weather data on a unified time axis, and generates prediction results accordingly. The data acquisition and prediction module acquires data (voltage, current, reactive power, power flow, circuit breaker status, etc.) from distribution transformers and their connected distribution feeders according to protocols such as IEC 61850, IEC 60870-5-104, or OPC UA through the interface between the station control layer and the field layer. It simultaneously acquires operating data from the distribution transformer and feeder sides, distribution automation terminals, smart meter curves, distributed power inverters, and energy storage converters, and integrates day-ahead and intraday electricity prices, as well as irradiance, temperature, wind speed, and wind direction from meteorological stations and numerical weather prediction.

[0011] The data acquisition and prediction module aligns and cleans the acquired data, then uses a prediction method based on historical time series and incorporating exogenous variables to generate prediction curves for future periods. For load and price series, it uses time series models or regression models with meteorological, holiday, and time-period dummy variables to fit each period and extrapolate on a rolling basis. For distributed power series, it uses meteorological elements as exogenous input to establish a regression / power curve model of the power-meteorological relationship and makes rolling predictions, thereby obtaining point prediction curves for each object in the future period. Based on this, the data acquisition and prediction module calculates the interval confidence boundary for each prediction curve according to the historical rolling residual distribution. Specifically, it statistically analyzes the residual sequence of point predictions within a recent window, estimates its upper / lower quantiles, and superimposes these quantile values ​​onto the point predictions according to time periods to form upper and lower bound trajectories; alternatively, it can directly output the upper / lower quantile prediction curves corresponding to the confidence level through quantile modeling.

[0012] As a further aspect of this invention, the power flow feasible region construction module generates a power flow feasible region based on the prediction curve and confidence interval, and outputs a time-based constraint set, including the following specific contents: The power flow feasible region construction module takes the prediction of each time point and its interval confidence boundary output by the data acquisition and prediction module as a premise, firstly, it performs scenario-based organization of the load, distributed power output and related meteorological drive within the future window on a unified time axis, and uniformly maps the upper and lower bounds of the injection power, the power factor range and the available interval of parallel compensation that can occur in each time period to the node level and feeder level of the distribution network; then, it loads the primary connection topology, line parameters, transformer capacity and node voltage limits, keeps the operation mode consistent, and evaluates the impact relationship of injection changes on node voltage and branch power flow in each time period within a given uncertainty interval, forming a feasible injection boundary corresponding to the upper / lower limit of voltage and the upper limit of line and transformer current carrying capacity consistent with the actual network boundary.

[0013] Based on this, the power factor index is transformed into an active-reactive coupling boundary, and the switching availability of the parallel compensation device during this period and the reactive power adjustment range are written together, so that reactive power and voltage constraints are consistently reflected in the same feasible domain. For reactive power settings of distributed generation sources, the power flow feasible region construction module limits its reactive power output range based on its grid connection capacity and power factor requirements, and merges it with the effective range of parallel compensation to avoid duplicate inclusion of reactive power contribution. For injection quantities with upper and lower bound prediction uncertainties, the power flow feasible region construction module takes the envelope according to the confidence boundary to ensure that the feasible region does not touch the voltage and current carrying limits under the most unfavorable injection disturbance, while setting a safety margin near the limits to buffer real-time deviations. When the feeder contains multiple distributed generation sources and load flexibility, the power flow feasible region construction module performs constraint superposition and transfer on the available active / reactive power ranges of each node, eliminates combinations that will cause node voltage over-limit or branch overload, and only retains the injection set that satisfies the upper and lower voltage limits, line and transformer capacity, power factor, and parallel compensation availability, thereby obtaining a power flow feasible region consistent with network status, equipment boundaries, and prediction uncertainties.

[0014] The power flow feasible region construction module dynamically updates the power flow feasible region according to a rolling step size. Each time the window moves forward, the boundary is recalculated based on the new point prediction and confidence boundary. This ensures that the multi-timescale rolling optimization module always solves within a safe region consistent with the current prediction. At the same time, the nodes or branches that are most likely to approach the upper limit are marked as "bottleneck positions" within the feasible region. This allows the optimization phase to identify and avoid solutions close to these boundaries outside of the cost objective. As a result, the "voltage upper and lower limits, line and transformer capacity, power factor and parallel compensation constraints" are completely and systematically transformed into a directly callable time-based constraint set, which serves as the input to the multi-timescale rolling optimization module.

[0015] As a further aspect of this invention, a multi-timescale rolling optimization module, within the feasible domain of power flow, establishes a time-segmented constraint set and solves it online using a feasible domain projection time-series scheduling algorithm. If the voltage proximity index exceeds a preset threshold, a short-window re-optimization is initiated to maintain power supply and demand balance. This includes the following specific steps: The multi-timescale rolling optimization module receives the constraint set provided by the feasible domain construction module by time period and simultaneously reads the prediction curve and confidence interval output by the data acquisition and prediction module, unifying them onto the time axis of the rolling window. For the aforementioned multi-source sequences, it performs missing measurement filling, anomaly truncation, clock alignment, and scale normalization, constructing a cross-window "state-disturbance-constraint" feature tensor, and generating multi-scale sample slices according to the rolling window hierarchy. Simultaneously, it generates compliance templates according to the node / feeder / equipment boundaries provided by the feasible domain of power flow. The multi-timescale rolling optimization module employs a feasible domain projection time-series scheduling algorithm to solve the problem online within the feasible domain of power flow, outputting the output, charging / discharging, and response plans for each unit in each time period to maintain power supply and demand balance. The feasible region projection time-series scheduling algorithm includes a time-series encoder, a constraint attention module, and a policy decoder. First, it reads the "point prediction and confidence boundary" and the power flow feasible region constraint set, and completes the alignment and normalization of sequences such as load, distributed power sources, price, and weather, constructing a multi-scale feature tensor covering "past state - current disturbance - future constraint". Then, the time-series encoder (a time convolutional network superimposed with bidirectional gated recurrent units) extracts cross-time period correlation and energy coupling clues. The constraint attention module applies higher weights to "bottlenecks" such as voltage-weak nodes, easily overloaded branches, and energy storage boundaries. The strategy decoder uses an autoregressive approach to progressively generate action vectors for each future time period, including distributed power active / reactive power settings, energy storage charging / discharging power and SOC trajectory, interruptible load shedding and restoration amounts, and parallel compensation switching suggestions. It also simultaneously provides upward / downward adjustments to reserve capacity and ramp-up estimates. The decoded actions are first mapped back into the power flow feasible domain via the power flow feasible domain projection layer. Then, the compliance template performs time-specific checks and fine-tuning of ramp-up, minimum start / stop, power factor, and reserve coverage, ensuring that the plan for each time period satisfies both instantaneous power balance and consistency of energy consumption and storage status across time periods. During online inference, the feasible domain projection timing scheduling algorithm generates the output, charging / discharging, and response plans for each unit within the entire window, along with their upper and lower quantile intervals, at the millisecond level. It also outputs proximity indicators for threshold triggering, forming timing settings that can be directly distributed to stations / feeders / equipment.

[0016] Based on the aforementioned timing settings, when the feasible region projection timing scheduling algorithm detects that the voltage proximity index exceeds a preset threshold during online monitoring, a short-window re-optimization process is initiated: First, executed instructions and current device status are frozen, and the latest measurements and recently corrected point predictions and confidence boundaries are uniformly aggregated to construct a shortened rolling window covering a few near-future time periods, using the original timing plan as the baseline trajectory; then, the feasible region constraint set of power flow is loaded one by one, while executability constraints and time coupling conditions are merged; in the solution phase, the feasible region projection timing scheduling algorithm first generates the corrected action vector within the short window, and any potential out-of-bounds solutions are immediately eliminated through the feasible region projection layer and compliance template; within the preset solution deadline of the short-window re-optimization, the lightweight numerical optimizer is called to optimize the solution. The above-mentioned correction scheme is refined locally, aiming to achieve rapid correction with the goals of minimizing the increase in overall cost, minimizing the penalty for breach of constraints, and minimizing the deviation from the baseline trajectory. If convergence is not achieved within the time limit, a machine learning backup scheme with projection verification is directly adopted to ensure executability. Finally, incremental setpoints and updated time series segments are generated and distributed hierarchically to the station level, feeder level, and equipment level. At the same time, the threshold table and backup arrangement are refreshed. Measurements and parameters are recorded throughout the process for writing back to the data acquisition and prediction module and the power flow feasible region construction module to continuously calibrate the model and boundaries. The loop is monitored to continue tracking proximity and error. If it recovers to within the threshold, the current short window optimization ends. Otherwise, it is triggered again according to the priority sequence to ensure that the operating point remains stably within the power flow feasible region and further maintain the coordinated balance of power and energy.

[0017] The lightweight numerical optimizer performs local refinement of the correction scheme as follows: After receiving the short-window correction scheme, the lightweight numerical optimizer uses the scheme as the initial value for hot start, limits the trust domain and step size, and searches only in the local neighborhood of key variables; the lightweight numerical optimizer simultaneously loads the power flow feasible domain constraints, voltage upper and lower limits, line and transformer capacity, power factor range, available range of parallel reactive power compensation, ramping and minimum start-stop, energy storage charging and discharging and state boundaries, and sets the objective as a weighted sum of minimizing the overall cost increment, minimizing constraint penalties, and minimizing deviation from the baseline trajectory; the solution uses a pre-compiled small-scale linear or second-order cone model, combined with sparse factor decomposition and early stop threshold to obtain fine-tuning in a very short time, refines the active and reactive power settings, charging and discharging power and switching actions, and performs a feasible domain projection check to ensure that the solution is within the safe domain. If the time budget is reached, the current optimal feasible solution is output and distributed.

[0018] The technical effects and advantages of this invention, a power supply and demand balance optimization system based on source-grid-load coordination, are as follows: This invention quantifies uncertainty and constructs a feasible power flow domain by using "predicted curves + confidence intervals." It unifies network physical boundaries such as voltage, current carrying capacity, power factor / reactive power compensation, and energy storage SOC into time-based constraints. Combining multi-timescale rolling optimization and "feasible domain projection" time-series scheduling, it triggers short-window re-optimization when voltage approaches or exceeds the threshold, achieving coordinated balance and rapid correction among source, grid, load, and storage. The resulting technical effects and advantages include: ① Improved safety—operating points remain within the feasible domain for extended periods, with the proportion of periods when voltage / current approaches or exceeds limits reduced to approximately 0.3% / 0.2%, respectively, and reserve gaps and ramp-up / minimum start-stop default events reduced to zero; ② Improved economic efficiency—curtailment rate reduced to approximately 2.1%, and cross-window power imbalance reduced to approximately 0.1%. MWh; ③ Enhanced executability – the planning curve is smoother and more traceable, and short window re-optimization reduces the end-to-end time of a single window to about 0.13s; ④ Improved asset utilization – the equivalent turnover of energy storage is increased to about 1.6 times / day; ⑤ Strong engineering adaptability – a closed loop of “collection – prediction – constraint mapping – optimization – correction” is formed, and bottleneck nodes / branches are automatically identified and avoided, significantly enhancing robustness and real-time performance compared to traditional single-time-scale static optimization. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of a power supply and demand balance optimization system based on source-grid-load coordination according to the present invention.

[0020] Figure 2 A flowchart illustrating the mathematical model for optimizing the dispatching of existing microgrid technologies.

[0021] Figure 3 This is a schematic diagram of the energy storage SOC and charge / discharge timing constraints of the present invention.

[0022] Figure 4 This is a schematic diagram of the feasible domain projection time-series scheduling algorithm of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1 like Figure 1As shown, the present invention provides a power supply and demand balance optimization system based on source-grid-load coordination, which includes: a data acquisition and prediction module, which acquires data of the distribution network and the external environment, and outputs prediction curves and confidence intervals; The power flow feasible region construction module generates the power flow feasible region based on the predicted curve and confidence interval, and outputs a time-based constraint set; The multi-timescale rolling optimization module, within the feasible domain of power flow, combines time-based constraint sets and solves the problem online using a feasible domain projection time-series scheduling algorithm. If the voltage proximity index exceeds a preset threshold, a short window is initiated for further optimization to maintain the balance between power supply and demand.

[0025] Furthermore, the data acquisition and prediction module is used to acquire and organize data such as distribution transformer / feeder measurements, distributed power generation, load, price, and weather on a unified time axis, and generate prediction results accordingly. The data acquisition and prediction module acquires data (voltage, current, reactive power, power flow, circuit breaker status, etc.) of distribution transformers and their connected distribution feeders through the station control layer and field layer interface, according to protocols such as IEC 61850, IEC 60870-5-104, or OPC UA. It simultaneously acquires operating data from the distribution transformer and feeder sides, distribution automation terminals, smart meter curves, distributed power inverters, and energy storage converters, and integrates day-ahead and intraday electricity prices, as well as irradiance, temperature, wind speed, and wind direction from meteorological stations and numerical weather prediction. The data acquisition and prediction module performs clock synchronization, missing value interpolation and outlier removal, noise reduction, and resampling on the acquired data, aligning second / minute and 15-minute / hourly data to a unified step size and completing node / feeder mapping.

[0026] The data acquisition and prediction module aligns and cleans the acquired data, and then uses a prediction method based on historical time series and incorporating exogenous variables to generate prediction curves for future periods: For load and price series, time series models (such as exponential smoothing / ARIMA family) or regression models with weather, holidays and time-period dummy variables are used to fit and extrapolate on a time-by-time basis; for distributed power series, a regression / power curve model of power-meteorological relationship is established with meteorological elements (such as irradiance, temperature, wind speed, etc.) as exogenous inputs and is predicted on a rolling basis, thereby obtaining point prediction curves for each object in future periods. Based on this, the data acquisition and prediction module calculates the interval confidence boundary for each prediction curve according to the historical rolling residual distribution: Specifically, it statistically analyzes the residual sequence of the predicted points within a recent window, estimates its upper / lower quantiles (such as the 5th / 95th quantile or the quantile that matches the target confidence level), and superimposes the quantile value onto the point prediction according to the time period to form the upper and lower bound trajectories; alternatively, it can directly output the upper / lower quantile prediction curves corresponding to the confidence level through quantile modeling (such as time series training of quantile regression or quantile loss).

[0027] Furthermore, the power flow feasible domain construction module, based on the time-point predictions and their interval confidence boundaries output by the data acquisition and prediction module, first performs scenario-based organization of the load, distributed power output, and related meteorological drivers within the future window on a unified time axis. It then maps the upper and lower bounds of the injection power, the power factor range, and the available intervals of parallel compensation for each time period to the node and feeder levels of the distribution network. Subsequently, it loads the primary wiring topology, line parameters, transformer capacity, and node voltage limits, maintaining consistent operation. For each time period, it evaluates the impact of injection changes on node voltage and branch power flow within a given uncertain interval, forming feasible injection boundaries corresponding to the upper / lower voltage limits and the upper current carrying limits of lines and transformers that are consistent with the actual network boundary.

[0028] The load within the future window refers to the time series forecast of electricity demand and its interval confidence boundaries for each node or feeder of the distribution network at the current moment, facing the forward time period (i.e., intraday-hourly / 15-minute-short-term / 5-minute-minute-level, etc., consistent with the system settings). The distributed power output refers to the actual or planned electrical power injected into the grid at the grid connection point by various distributed generation units connected to the distribution network, such as photovoltaic, wind power, distributed natural gas units, fuel cells, small hydropower, etc. The distributed power output includes active power output and reactive power output. The active power output refers to the amount of electrical power provided to the grid at a certain moment / period, and the reactive power output represents the reactive power provided or absorbed by the distributed power source at the grid connection point. The meteorological drive refers to the meteorological elements that directly or indirectly determine the changes in distributed power output and electricity load. They are incorporated into the data acquisition and forecasting module as exogenous inputs to generate the forecast curve for future periods and its upper and lower fluctuation ranges.

[0029] Based on this, the power factor index is transformed into an active-reactive coupling boundary, and the switching availability of parallel compensation devices (such as capacitor banks and reactive power compensation devices) during this period and the reactive power adjustment range are written together, so that reactive power and voltage constraints are consistently reflected in the same feasible domain. For reactive power settings of distributed generation sources, the power flow feasible region construction module limits its reactive power output range based on its grid connection capacity and power factor requirements, and merges it with the effective range of parallel compensation to avoid duplicate inclusion of reactive power contribution. For injection quantities with upper and lower bound prediction uncertainties, the power flow feasible region construction module takes the envelope according to the confidence boundary to ensure that the feasible region does not touch the voltage and current carrying limits under the most unfavorable injection disturbance, while setting a safety margin near the limits to buffer real-time deviations. When the feeder contains multiple distributed generation sources and load flexibility, the power flow feasible region construction module performs constraint superposition and transfer on the available active / reactive power ranges of each node, eliminates combinations that will cause node voltage over-limit or branch overload, and only retains the injection set that satisfies the upper and lower voltage limits, line and transformer capacity, power factor, and parallel compensation availability, thereby obtaining a power flow feasible region consistent with network status, equipment boundaries, and prediction uncertainties.

[0030] The feasible power flow domain is based on the predictions for each time period and their upper and lower confidence boundaries given by the data acquisition and prediction module. It maps the possible source, load, and storage injection ranges to each node / feeder. Within the network physical boundary, it filters out combinations that would lead to voltage overruns or branch overloads, and what remains is the feasible domain for that time period. At the same time, it provides quantitative indicators such as "proximity" to describe the safety margin of the operating point from the boundary.

[0031] The power flow feasible region construction module dynamically updates the power flow feasible region according to a rolling step size. Each time the window moves forward, the boundary is recalculated based on the new point prediction and confidence boundary. This ensures that the multi-timescale rolling optimization module always solves within a safe region consistent with the current prediction. At the same time, the nodes or branches that are most likely to approach the upper limit are marked as "bottleneck positions" within the feasible region. This allows the optimization phase to identify and avoid solutions close to these boundaries outside of the cost objective. As a result, the "voltage upper and lower limits, line and transformer capacity, power factor and parallel compensation constraints" are completely and systematically transformed into a directly callable time-based constraint set, which serves as the input to the multi-timescale rolling optimization module.

[0032] Furthermore, the multi-time-scale rolling optimization module receives the constraint set provided by the power flow feasible region construction module according to time periods, and synchronously reads the prediction curve and confidence interval output by the data acquisition and prediction module, unifying them onto the time axis of the rolling window. Table 1 shows the basic operating conditions of the multi-time-scale rolling optimization module in this embodiment.

[0033] Table 1 Experimental conditions and data sources

[0034] For the aforementioned multi-source sequences, missing measurement imputation, anomaly truncation, clock alignment, and scale normalization are performed. A cross-window "state-perturbation-constraint" feature tensor is constructed, and multi-scale sample slices are generated according to the rolling window level (e.g., intraday-hour level, 15-minute-short-term level, and 5-minute-minute level). At the same time, a "compliance template" is generated according to the node / feeder / equipment boundaries provided by the power flow feasible domain. The compliance template is a set of machine-readable constraints and mappings generated by the power flow feasible domain construction module in each rolling period and can be directly called by the multi-timescale rolling optimization module.

[0035] The multi-timescale rolling optimization module employs a feasible region projection time-series scheduling algorithm. Within the feasible region of power flow, it outputs the output, charging / discharging, and response plans of each unit at each time period through online rolling solutions, maintaining a balance between power supply and demand. Figure 3As shown, the feasible region projection time-series scheduling algorithm includes a time-series encoder, a constraint attention module, and a policy decoder. First, it reads the "point prediction and confidence boundary" and the feasible region constraint set for power flow, aligning and normalizing sequences such as load, distributed generation, price, and weather, constructing a multi-scale feature tensor covering "past state—current disturbance—future constraint." Then, the time-series encoder (a time convolutional network superimposed with bidirectional gated recurrent units) extracts cross-time-period correlation and energy coupling clues. The constraint attention module applies higher weights to "bottlenecks" such as voltage-weak nodes, easily overloaded branches, and energy storage boundaries, enabling the feasible region projection time-series scheduling algorithm to prioritize these critical areas—voltage-weak, easily overloaded, and SOC boundaries—when generating plans, thus mitigating problems early, minimizing boundary touches, and reducing passive corrections, thereby simultaneously improving the operational safety and executability of distribution network equipment. The strategy decoder uses an autoregressive approach to progressively generate action vectors for each future time period, including active / reactive power settings for distributed power sources, energy storage charging and discharging power and SOC trajectory, interruptible load shedding and restoration amounts, and parallel compensation switching suggestions. It also simultaneously provides upward / downward adjustments to reserve margins and ramp-up estimates. The decoded actions are first mapped back into the feasible power flow domain through the power flow feasible domain projection layer, and then the compliance template performs time-specific checks and fine-tuning of ramp-up, minimum start / stop, power factor, and reserve coverage to ensure that the plan for each time period satisfies both instantaneous power balance and consistency of energy consumption and energy storage status across time periods. During the offline training phase, the feasible domain projection time-series scheduling algorithm uses historical operating data and near-optimal solutions obtained through offline optimization based on feasible domain and time-specific constraints as supervision signals to construct a comprehensive objective function. Firstly, it incorporates economic considerations reflected in electricity purchase costs, curtailment penalties, interruptible load compensation expenditures, and energy storage lifetime depreciation costs. Secondly, it adds a penalty term for constraint violation behavior to prevent the breach of conditions such as voltage, power flow, capacity, power factor, and reactive power compensation. Thirdly, it introduces penalty terms for the smoothness of the planning curve and the executability of ramping, to suppress unrealistic jumps between adjacent time periods and ensure on-site traceability. Fourthly, it uses quantile loss to characterize the upper and lower bounds of the prediction interval, thereby calibrating the upper and lower boundaries of the output plan under uncertain conditions. By minimizing this multi-factor objective, the model simultaneously learns preferences for economic optimality, network security, and executability during training, forming a joint decision criterion of "cost-security-executability," and laying the foundation for rapidly generating output, charging, discharging, and response plans for each unit in each time period within the feasible power flow domain during online inference.When the feasible domain projection timing scheduling algorithm is in online inference, it generates, within milliseconds, the output (active and reactive power output of distributed power sources), charging and discharging (charging power / discharging power of energy storage units in each time period), and response plans (for interruptible loads, parallel reactive power compensation devices, and up / down reserve resources, the timing arrangement of their cut-off or restoration amounts, reactive power switching / setting, start-up time, duration, and recovery slope is arranged according to time periods, which is used to quickly intervene when threshold triggering or disturbance occurs to maintain voltage and power flow within the limits and pull the operating point back into the feasible domain of power flow) and their upper and lower quantile intervals, and outputs the proximity index (a measure of the remaining margin of the operating point from various constraint boundaries such as voltage, power flow, SOC, and reserve within the feasible domain of power flow using a unified dimension) for threshold triggering, forming timing setting values ​​that can be directly issued to the station / feeder / equipment.

[0036] Based on the aforementioned timing settings, when the feasible region projection timing scheduling algorithm detects that the voltage proximity index exceeds a preset threshold during online monitoring, a short-window re-optimization process is initiated: First, the executed instructions and current equipment status are frozen, and the latest measurements and recently corrected point predictions and confidence boundaries are uniformly aggregated to construct a shortened rolling window covering a few near-future time periods (i.e., 5-10 minute time periods), using the original timing plan as the baseline trajectory; then, the feasible region constraint set of power flow is loaded one by one, while merging executability constraints and time coupling conditions, such as ramp limits, minimum start / stop duration, backup coverage, and upper limits for energy storage charging and discharging (e.g., Figure 4 As shown in the figure, the state boundary is defined, and the executed quantities and triggering order are set as hard or high-weight soft constraints; in the solution phase, the feasible region projection timing scheduling algorithm first generates the corrected action vector within the short window, and any potential out-of-bounds solutions are immediately eliminated through the feasible region projection layer and compliance template; within the preset solution deadline of the short window re-optimization, the lightweight numerical optimizer is called to locally refine the above correction scheme, with the goal of minimizing the incremental overall cost, minimizing the constraint violation penalty, and minimizing the deviation from the baseline trajectory to complete the fast correction. If it does not converge within the time limit, the projected correction is directly adopted. The core machine learning backup scheme ensures execution; ultimately, incremental setpoints and updated time series segments are generated and distributed hierarchically to the station level, feeder level, and equipment level. At the same time, the threshold table and backup orchestration are refreshed. Measurements and parameters are recorded throughout the process for writing back to the data acquisition and prediction module and the power flow feasible region construction module to continuously calibrate the model and boundaries. The loop is monitored to continue tracking proximity and error. If it recovers to within the threshold, the current short window optimization ends. Otherwise, it is triggered again according to the priority sequence to ensure that the operating point remains stably within the power flow feasible region and further maintain the coordinated balance of power and energy.

[0037] In this embodiment, as shown in Table 2, the core indicators of the complete 24-hour simulation are compared. The three methods are compared under the same baseline: traditional single-timescale static optimization, numerical optimization without machine learning but internalizing the feasible region, and feasible region projection time-series scheduling plus short window re-optimization of the present invention.

[0038] Table 2 Comparison of 24-hour main indicators

[0039] As can be seen from Table 2, under the conditions of equivalent network and prediction error, after internalizing the feasible region of power flow and introducing feasible region projection time-series scheduling, the proportion of time periods when voltage and current are close to or exceed the limit decreases significantly, and the curtailment rate and cross-window power imbalance decrease simultaneously. This proves that under the same constraints and uncertainties, this scheme can simultaneously reduce the exceedance and curtailment, converge energy balance, reduce default and shorten the time delay.

[0040] The lightweight numerical optimizer performs local refinement of the correction scheme as follows: After receiving the short-window correction scheme, the lightweight numerical optimizer uses the scheme as the initial value for hot start, limits the trust domain and step size, and searches only in the local neighborhood of key variables; the lightweight numerical optimizer simultaneously loads the power flow feasible domain constraints, voltage upper and lower limits, line and transformer capacity, power factor range, available range of parallel reactive power compensation, ramping and minimum start-stop, energy storage charging and discharging and state boundaries, and sets the objective as a weighted sum of minimizing the overall cost increment, minimizing constraint penalties, and minimizing deviation from the baseline trajectory; the solution uses a pre-compiled small-scale linear or second-order cone model, combined with sparse factor decomposition and early stop threshold to obtain fine-tuning in a very short time, refines the active and reactive power settings, charging and discharging power and switching actions, and performs a feasible domain projection check to ensure that the solution is within the safe domain. If the time budget is reached, the current optimal feasible solution is output and distributed.

[0041] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0042] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power supply and demand balance optimization system based on source-grid-load coordination, characterized in that, include: The data acquisition and prediction module acquires data from the power distribution network and the external environment, and outputs prediction curves and confidence intervals. The power flow feasible region construction module generates the power flow feasible region based on the predicted curve and confidence interval, and outputs a time-based constraint set; The multi-timescale rolling optimization module, within the feasible domain of power flow, combines time-based constraint sets and solves the problem online based on the feasible domain projection time-series scheduling algorithm. If the voltage proximity index exceeds the preset threshold, it initiates short-window re-optimization to maintain the balance between power supply and demand. The feasible region projection time-series scheduling algorithm includes a time-series encoder, a constraint attention module, and a policy decoder. First, it reads the point prediction, confidence boundary, and power flow feasible region constraint set to construct a multi-scale feature tensor. Then, it extracts cross-time-period correlation and energy coupling clues. Next, it applies higher weights to voltage-weak nodes, easily overloaded branches, and energy storage boundaries to improve the operational safety of distribution network equipment. During online inference, the algorithm generates the output, charging / discharging, and response plans for each unit within the entire window, along with their upper and lower intervals, at the millisecond level. It also outputs a proximity index for threshold triggering, forming a time-series setpoint. Based on this setpoint, when the algorithm detects that the voltage proximity index exceeds a preset threshold during online monitoring, it initiates a short-window re-optimization process to keep the operating point stable within the feasible region and further maintain the coordinated balance between power and energy consumption.

2. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 1, characterized in that... The strategy decoder uses an autoregressive approach to progressively generate action vectors for each future time period, including active / reactive power settings for distributed power sources, energy storage charging and discharging power and SOC trajectory, interruptible load shedding and restoration amounts, and parallel compensation switching suggestions. It also provides up / down reserve margin and ramp-up estimates. The decoded actions are first mapped back into the power flow feasible domain through the power flow feasible domain projection layer, and then the compliance template performs time-based verification and fine-tuning to ensure that the plan for each time period satisfies both instantaneous power balance and consistency of power and energy storage status across time periods.

3. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 2, characterized in that... The short window re-optimization first freezes the executed instructions and the current device status, unifies the latest measurements and point predictions and confidence boundaries based on recent corrections, constructs a shortened rolling window, and uses the original time-series plan as the baseline trajectory. Then, the set of feasible domain constraints for power flow is loaded one by one, while executability constraints and time coupling conditions are merged. During the solution phase, the feasible region projection timing scheduling algorithm first generates the correction action vector within the short window, forming incremental setpoints and updated timing segments, which are then distributed hierarchically to the station level, feeder level, and equipment level. The threshold table and backup arrangement are refreshed, and measurements and parameters are recorded throughout the process. These are used to write back to the data acquisition and prediction module and the power flow feasible region construction module to continuously calibrate the model and boundaries. The loop is monitored to continue tracking proximity and error. If the loop recovers to within the threshold, the current round of short window optimization ends; otherwise, it is triggered again according to the priority sequence.

4. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 3, characterized in that, Within the preset solution deadline of the short-window re-optimization, the lightweight numerical optimizer is invoked to locally refine the correction scheme, aiming to achieve rapid correction with the goals of minimizing the overall cost increment, minimizing the constraint violation penalty, and minimizing the deviation from the baseline trajectory.

5. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 1, characterized in that, The data acquisition and prediction module acquires data from distribution transformers, feeders, distributed power inverters, energy storage converters, distribution automation terminals, and smart meters via IEC 61850, IEC 60870-5-104, or OPC UA protocols, and integrates time-of-use pricing and meteorological data. It performs clock synchronization, missing data interpolation, anomaly removal, noise reduction, and resampling, aligning second-level or minute-level data with 15-minute or hour-level data to a uniform step size and completing node and feeder mapping.

6. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 1, characterized in that, The power flow feasible domain construction module uniformly maps the upper and lower bounds of injected power, the power factor range, and the available interval of parallel reactive power compensation devices to the node and feeder levels. It loads the primary wiring topology, line parameters, transformer capacity, and node voltage limits to evaluate the impact of injection changes on voltage and power flow. It sets reactive power coupling boundaries and configures safety margins near the limits. It identifies nodes or branches that are close to the upper limit as bottleneck locations and outputs machine-readable time-based constraint sets.

7. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 1, characterized in that, The time-series encoder is composed of a time convolutional network and a bidirectional gated recurrent unit to extract cross-time-period correlation and energy coupling clues. The constrained attention module assigns higher weights to voltage-weak nodes, easily overloaded branches and energy storage boundaries.

8. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 2, characterized in that, The compliance template is generated by the power flow feasible domain construction module in each rolling period and includes the mapping relationship between nodes, feeders and equipment boundaries, power factor coupling boundaries, parallel compensation available range and safety margin parameters. It is directly called by the multi-time scale rolling optimization module to ensure that the plans for each period simultaneously meet the instantaneous power balance and cross-period power consistency.

9. The power supply and demand balance optimization system based on source-grid-load coordination according to claim 1, characterized in that, The multi-timescale rolling optimization adopts a main step size and multi-layer forward window rolling mechanism. The main step size is 15 minutes, and the forward window includes a 24-hour level, a 1-hour level, and a 5-minute short window level. For different levels, a multi-scale feature tensor of "state, perturbation and constraint" across the window is constructed and corresponding sample slices are generated.

Images