Intelligent low-voltage flexible direct current environment self-adapting multi-element energy utilization control method

Abstract

The application relates to an intelligent low-voltage flexible direct current environment self-adaptive multi-element energy utilization control method. The method comprises the following steps: acquiring power system distribution network load data, distribution network line data and distribution network environment data; performing dynamic energy utilization demand adaptation analysis on the power system according to the distribution network load data, the distribution network line data and the distribution network environment data to obtain energy utilization demand adaptation analysis data; performing dynamic energy utilization demand adaptation mapping on the power system according to the energy utilization demand adaptation analysis data to obtain energy utilization demand adaptation mapping instructions; controlling the power system to perform electric energy state conversion according to the energy utilization demand adaptation mapping instructions to obtain distribution network operation state data; and optimizing and adjusting the operation strategy of the power system according to the distribution network operation state data to obtain distribution network energy utilization dynamic control data. The method can simultaneously solve the composite problem of the low-voltage power supply scheme.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, and in particular to a smart low-voltage flexible DC-environment adaptive multi-electrode energy control method, device and computer equipment. Background Technology

[0002] Traditional technologies can reduce line loss to 10% by deploying multiple reactive power compensation nodes in the line, but they do not solve the problem of voltage drop over long distances. Low-temperature adaptable flexible DC modules use wide-temperature-range IGBT devices, achieving a 95% start-up success rate at -20℃, but lack a coordinated control mechanism with reactive power compensation. The NB-IoT smart monitoring system piloted by local power companies can remotely collect load data, but it cannot coordinate with the power conversion module for control. Therefore, traditional technologies cannot simultaneously solve the complex problems of low-voltage power supply solutions. Summary of the Invention

[0003] Therefore, it is necessary to provide an intelligent low-voltage flexible DC-environment adaptive multi-electrode energy control method, device, and computer equipment that can simultaneously solve the complex problems of low-voltage power supply schemes, addressing the aforementioned technical issues.

[0004] In a first aspect, this application provides an intelligent low-voltage flexible DC-environment adaptive multi-element energy control method, including:

[0005] Acquire power system distribution network load data, distribution network line data, and distribution network environment data;

[0006] Based on the distribution network load data, the distribution network line data, and the distribution network environment data, a dynamic energy demand adaptation analysis is performed on the power system to obtain energy demand adaptation analysis data.

[0007] Based on the energy demand adaptation analysis data, the power system is dynamically mapped to meet energy demand, and an energy demand adaptation mapping instruction is obtained.

[0008] According to the energy demand adaptation mapping instruction, control the power system to perform power state conversion and obtain distribution network operation status data;

[0009] Based on the distribution network operation status data, the operation strategy of the power system is optimized and adjusted to obtain dynamic control data of distribution network energy consumption.

[0010] Secondly, this application also provides an intelligent low-voltage flexible DC-environment adaptive multi-energy control device, comprising:

[0011] The distribution network data acquisition module is used to acquire distribution network load data, distribution network line data, and distribution network environment data of the power system.

[0012] The energy demand adaptation module is used to perform dynamic energy demand adaptation analysis on the power system based on the distribution network load data, the distribution network line data and the distribution network environment data, and obtain energy demand adaptation analysis data.

[0013] The energy demand adaptation module is also used to perform dynamic energy demand adaptation mapping on the power system based on the energy demand adaptation analysis data, and obtain energy demand adaptation mapping instructions.

[0014] The power state conversion module is used to adapt the mapping instruction according to the energy demand, control the power system to perform power state conversion, and obtain distribution network operation status data.

[0015] The operation strategy optimization module is used to optimize and adjust the operation strategy of the power system based on the distribution network operation status data to obtain dynamic control data of distribution network energy consumption.

[0016] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of an intelligent low-voltage flexible DC-environment adaptive multi-electrode energy control.

[0017] The aforementioned intelligent low-voltage flexible DC-environment adaptive multi-source energy control method, device, and computer equipment, through the fusion of multi-source data from distribution network load, lines, and environment, first comprehensively grasps the linkage between equipment health, power flow distribution, and external disturbances (meteorology, temperature and humidity, disaster risks, etc.) from a spatiotemporal perspective. Then, through dynamic energy demand adaptation analysis, it accurately quantifies user-side energy preferences and grid-side constraints, and subsequently maps the analysis results into directly executable power state transition and dispatch commands. This enables the distribution network to quickly and finely complete power flow reconstruction and power allocation under different load peaks and valleys, meteorological changes, or abnormal operating conditions. It can simultaneously solve the complex problems of long-distance transmission efficiency, dynamic reactive power compensation, and extreme environmental tolerance, while reducing operation and maintenance difficulty and costs, and supporting remote monitoring and fault self-diagnosis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an application environment diagram of the intelligent low-voltage flexible DC-environment adaptive multi-electrode energy control method in one embodiment;

[0020] Figure 2This is a flowchart illustrating an intelligent low-voltage flexible DC-environment adaptive multi-electrode energy control method in one embodiment;

[0021] Figure 3 This is a structural block diagram of an intelligent low-voltage flexible DC-environment adaptive multi-energy control device in one embodiment;

[0022] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] This application provides an intelligent low-voltage flexible DC-environment adaptive multi-element energy control method, which can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0025] In one exemplary embodiment, such as Figure 2 As shown, an intelligent low-voltage flexible DC-environment adaptive multi-element energy control method is provided, which is applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 210. Wherein:

[0026] Step 202: Obtain power system distribution network load data, distribution network line data, and distribution network environment data.

[0027] Step 204: Based on the distribution network load data, distribution network line data, and distribution network environment data, perform dynamic energy demand adaptation analysis on the power system to obtain energy demand adaptation analysis data.

[0028] Step 206: Based on the energy demand adaptation analysis data, perform dynamic energy demand adaptation mapping on the power system to obtain energy demand adaptation mapping instructions.

[0029] Step 208: Based on the energy demand adaptation mapping instruction, control the power system to perform power state transition and obtain distribution network operation status data.

[0030] Step 210: Based on the distribution network operation status data, optimize and adjust the power system operation strategy to obtain distribution network energy consumption dynamic control data.

[0031] Among them, distribution network load data is time-series measurement data that reflects the behavior of the electricity consumption side, including three-phase voltage, current, active / reactive power, power factor and harmonics, etc., which are used to characterize load characteristics and fluctuations.

[0032] Among them, distribution network line data are parameters and measurements that characterize the transmission and distribution status of distribution lines, including line topology, conductor and cable parameters, voltage and current at the beginning and end, equivalent impedance and temperature along the line.

[0033] Among them, the distribution network environment data are the external and internal environmental quantities that affect the operation of the distribution network, such as temperature, humidity, dust concentration, altitude / air pressure, wind speed and equipment cavity temperature rise.

[0034] Dynamic energy demand adaptation analysis is a process of jointly assessing and classifying energy demand based on load, line and environmental data, focusing on targets such as voltage quality, line loss and harmonics.

[0035] Among them, the energy demand adaptation analysis data is a structured result set formed by the above analysis process, which includes information such as scenario tags, priorities, target weights, and constraints / thresholds.

[0036] Among them, dynamic energy demand adaptation mapping is the process of transforming analytical data into executable control strategies, including scenario-based parameter solving and cross-scenario conflict resolution.

[0037] Among them, the energy demand adaptation mapping instruction is a set of control instructions generated and issued during the mapping process, such as the DC side target voltage Udc*, the reactive power instructions Q_i* for each segment of SVG, the inverter voltage regulation / harmonic suppression parameters, and the environmental action threshold.

[0038] Among them, the power state transition is a comprehensive control action by which the execution layer, based on instructions, coordinates rectification, reactive power compensation, inverter and environmental devices to bring the power state into the target range.

[0039] Among them, the distribution network operation status data is a real-time operation data packet that is collected and encapsulated after the power state transition, including voltage / current / power factor, harmonics, line loss, temperature rise, actuator status and alarm information.

[0040] Among them, the operating strategy is the overall control rule formulated by the master controller for the current and predicted operating conditions, including control gain, target setting, and execution sequence.

[0041] Among them, the dynamic control data of power distribution network is a set of control quantities and parameters for the current cycle obtained by optimization / predictive control solution, which is used to drive the rectified voltage, reactive power distribution, inverter settings and environmental actions in the next cycle.

[0042] Specifically, on the edge controller (PLC / industrial computer), data on distribution network load, distribution network lines, and distribution network environment are synchronously collected at a period of ≤100 ms. For example, load-side data includes three-phase voltage, current, active / reactive power, power factor, and harmonic currents (2nd–31st order); line-side data includes voltage / current at both ends and fiber optic distributed temperature measurement (DTS) / conductor online temperature; and environmental data includes cabinet interior and exterior temperature, relative humidity, dust concentration, altitude / air pressure, and wind speed. The sampled data is Kalman filtered and 3σ anomaly removed, then encapsulated into timestamp frames and transmitted back to the server and cloud via NB-IoT / RS-485 dual-channel for backup.

[0043] The server processes three types of data in parallel: distribution network load data, distribution network line data, and distribution network environment data. First, it performs wavelet feature extraction and power change rate ΔP / Δt evaluation on the load data, obtaining labels such as "steady-state / abrupt / impact". Second, it calculates voltage drops and line losses along the line based on temperature-corrected impedance for the line data. Third, it maps the environmental data into environmental impact coefficients (low temperature / high dust / high altitude reduction tables). Then, using "minimum voltage deviation + minimum line loss + THD constraint" as multiple objectives, it performs a joint evaluation of the above three types of features to form energy demand adaptation analysis data (including demand level, priority, and constraint set).

[0044] Energy demand adaptation analysis data is decomposed into multiple scenarios (e.g., high-power motor startup, periodic constant-temperature load, peak-valley overlap for residential use, sandstorm / low-temperature extreme environments, etc.), and energy consumption parameter strategies are generated for each scenario (DC-side target Udc*, SVG segmented reactive power command Qi*, inverter voltage regulation / tuning parameters, environmental execution thresholds). Conflict detection and trade-offs are performed on the strategies for each scenario (voltage quality takes priority, followed by line loss and switching costs), and energy demand adaptation mapping commands (including timeout and rollback strategies) are formed through weighted fusion mapping.

[0045] The server further adapts and links with mapping instructions according to energy demand. On the rectifier side, adaptive PI / fuzzy compensation is used to increase or decrease Udc to offset line voltage drop. On the line side, each segment of SVG distributes reactive power according to "impedance ratio + current sharing constraint" and quickly suppresses voltage fluctuations. On the inverter side, V / I dual-loop + SVPWM (including three-harmonic injection) stabilizes the voltage to the target range and APF compensates for harmonics in real time. The environmental module triggers low-temperature heating / filter self-cleaning / air duct switching according to thresholds. The execution results are collected to form distribution network operation status data (voltage / harmonics / line loss / temperature rise / actuator status and alarms).

[0046] Within the rolling window, the execution deviation between the distribution network operation status data and the energy demand adaptation mapping command is input into the differentiable robust MPC. The Wasserstein spherical uncertainty set is used to characterize load and line disturbances, constructing a prediction cost that includes voltage deviation, reactive power output upper limit / slope, switching penalties, and temperature rise constraints. Then, the prediction model is quadratic programming (QP) and implicitly parameterized to allow the solver to adaptively tune based on historical deviations and environmental impact coefficients. Finally, within each control step, the dynamic control data of distribution network energy consumption (including Udc*, Qi* of each segment's SVG, inverter voltage regulation / tuning parameters, and environmental threshold adjustments) are solved, and dead zones, rate of change limits, and minimum dwell times are applied simultaneously to suppress frequent switching.

[0047] In the aforementioned intelligent low-voltage flexible DC-environment adaptive multi-source energy control method, multi-source data fusion of distribution network load, lines, and environment is used to first comprehensively grasp the linkage between equipment health, power flow distribution, and external disturbances (meteorology, temperature and humidity, disaster risks, etc.) from a spatiotemporal perspective. Then, through dynamic energy demand adaptation analysis, user-side energy preferences and grid-side constraints are accurately quantified. Subsequently, the analysis results are mapped into directly executable power state transition and dispatch commands, enabling the distribution network to quickly and finely complete power flow reconstruction and power allocation under different load peaks and valleys, meteorological changes, or abnormal operating conditions. It can simultaneously solve the complex problems of long-distance transmission efficiency, dynamic reactive power compensation, and extreme environmental tolerance, while reducing operation and maintenance difficulty and costs, and supporting remote monitoring and fault self-diagnosis.

[0048] In an exemplary embodiment, dynamic energy demand adaptation analysis is performed on the power system based on distribution network load data, distribution network line data, and distribution network environment data to obtain energy demand adaptation analysis data, including steps 302 to 308. Wherein:

[0049] Step 302: Perform load characteristic calculations on the distribution network load data to obtain distribution network load characteristic information;

[0050] Step 304: Perform line parameter calculations on the distribution network line data to obtain distribution network line characteristic information;

[0051] Step 306: Perform environmental impact calculations on the distribution network environment data to obtain distribution network environmental characteristic information;

[0052] Step 308: Based on the distribution network load characteristic information, distribution network line characteristic information, and distribution network environment characteristic information, perform dynamic energy demand adaptation analysis on the power system to obtain energy demand adaptation analysis data.

[0053] Among them, load characteristic calculation is the calculation and processing of distribution network load data by denoising and feature extraction (such as power change rate, power factor, THD, three-phase imbalance and frequency band energy).

[0054] Among them, the distribution network load characteristic information is a structured result obtained from load characteristic calculations, which summarizes the load's steady-state / abrupt label, PF, THD, ΔP / Δt, unbalance degree, and typical spectral characteristics.

[0055] Among them, the line parameter calculation is the process of identifying and calculating the line resistance / reactance / impedance, voltage drop, line loss, and reactive power sensitivity based on the measurement and temperature data at the beginning and end of the line.

[0056] Among them, the distribution network line characteristic information is a structured result formed by the calculation of line parameters, including R, X, equivalent Z, voltage drop / line loss distribution, bottleneck nodes and segmented reactive power support margin, etc.

[0057] Among them, the environmental impact calculation is the calculation process that maps environmental quantities such as temperature, humidity, dust, and altitude to equipment derating, ventilation efficiency and operational risk coefficient.

[0058] Among them, the distribution network environmental characteristic information is a structured result obtained from environmental impact calculation, which provides reduction coefficients such as temperature / altitude / dust and comprehensive environmental impact coefficient and related action thresholds.

[0059] Specifically, the edge controller reads measurements such as three-phase voltage / current / power factor / harmonic current at a period of ≤100 ms. After denoising the original waveform of the distribution network load data (Kalman or low-pass + 3σ rejection), it calculates active / reactive power, ΔP / Δt (power change rate), unbalance, and voltage / current THD. It also performs wavelet energy feature analysis and short-window FFT on the current and voltage sequences within a 1 s sliding window to identify load types such as impact / periodic / steady-state loads. The output of distribution network load characteristic information includes structured fields such as {steady-state / abrupt label, ΔP / Δt threshold flag, PF, THD, three-phase unbalance, typical frequency band energy, and power consumption scenario label}.

[0060] Based on the voltage and current at the beginning and end of the distribution network line data, the segmented current, and the DTS / temperature measurement point data along the line, first determine the resistance R(T) = R(T) based on the conductor temperature. 20 The voltage drop curve along the line is calculated as follows: [1+α(T−20℃)] is corrected, and the reactance X is estimated by combining the line length and cross-section. The impedance identification (Z=ΔU / ΔI) and topology consistency verification are performed in the small waveband of the operating condition using synchronous measurement. The voltage drop curve, segmented line loss, bottleneck node and available reactive power support margin along the line are calculated to form the characteristic information of the distribution network line, including {R, X, equivalent Z, voltage drop / line loss distribution, key node number, reactive power sensitivity coefficient of each segment}, etc.

[0061] Based on environmental data such as temperature, humidity, dust concentration, altitude / air pressure, and wind speed inside and outside the distribution network cabinet, the rated current derating factors k_T and k_H of the devices are calculated according to the insulation and heat dissipation derating curves for temperature and altitude. The ventilation efficiency factor k_D is estimated based on the filter pressure difference and air intake cleanliness for dust. Creepage / condensation risk indicators are given for high humidity environments. The multi-factor environmental impact coefficient K_env (0–1) is synthesized through normalization and rule base / simplified regression model, and the distribution network environmental characteristic information is output, including {k_T, k_H, k_D, K_env, environmental action threshold and priority to be triggered}.

[0062] After normalizing the distribution network load characteristics, distribution network line characteristics, and distribution network environmental characteristics according to the same dimensions, a multi-objective evaluation model is constructed with the minimum terminal voltage deviation as the primary objective and the minimum line loss and THD ≤ threshold as secondary objectives. The constraint set includes the upper limit of line reactive power capacity / change penalty, Udc safety boundary, and the output range caused by environmental reduction K_env. The objective weights are calculated and feasible regions are determined for scenarios such as "steady state / impact / extreme environment". The demand level and priority, key controlled quantities (Udc*, sensitivity weight of segmented Qi), and allowable switching frequency are given for each scenario, and summarized into energy demand adaptation analysis data.

[0063] In this embodiment, by performing characteristic / parameter / impact calculations on three types of data—load, line, and environment—and jointly analyzing them in a unified evaluation model, a high-precision characterization and differentiated weight allocation of the actual operating conditions can be achieved. This results in a significant reduction in voltage deviation and line loss, effective suppression of harmonics, and more accurate setting of key parameters such as reactive power compensation and DC bus voltage, which better align with the current scenario and safety boundaries. This improves the targeting and response speed of the strategy, reduces unnecessary switching and setting times, and avoids overload and instability caused by external factors such as temperature, dust, and altitude. Ultimately, under the same power supply capacity, better energy efficiency indicators, equipment lifespan, and power supply reliability are obtained.

[0064] In an exemplary embodiment, dynamic energy demand adaptation analysis is performed on the power system based on distribution network load characteristic information, distribution network line characteristic information, and distribution network load characteristic information to obtain energy demand adaptation analysis data, including steps 402 to 408. Wherein:

[0065] Step 402: Based on the distribution network environment characteristic information, perform counterfactual causal correction on the distribution network load characteristic information to obtain load counterfactual characteristic information;

[0066] Step 404: Perform physical constraint graph homotopy continuation processing on the impedance and topology data in the distribution network line characteristic information to obtain the line load coupling manifold parameters;

[0067] Step 406: Based on the load counterfactual characteristic information and the line load coupling manifold parameters, perform weighted least squares evaluation on each power parameter of the power system to obtain the target evaluation vector;

[0068] Step 408: Based on the target evaluation vector, perform dual primal co-solution of differentiable mixed integer programming on the power system to obtain energy demand adaptation analysis data.

[0069] Among them, counterfactual causal correction is to use causal methods to eliminate the influence of environmental and other confounding factors on load characteristics while maintaining the physical meaning, and to infer the "de-biased" load characterization that should be presented under the baseline environment.

[0070] Among them, the load counterfactual characteristic information is a set of load characteristic results obtained after counterfactual causal correction, reflecting key quantities such as power factor, fluctuation degree, harmonics and imbalance degree under baseline environmental conditions.

[0071] Impedance and topology data are a set of parameters describing the distribution network structure and electrical connections, including node-branch relationships, switch status, line resistance and reactance, and equivalent impedance.

[0072] Among them, the physical constraint graph homotopy continuation processing is a graphical model solution method that, under the premise of satisfying power flow and equipment boundary conditions, gradually and continuously transforms the simplified network into a real network and simultaneously obtains the system response and sensitivity.

[0073] Among them, the line load coupling manifold parameter is a low-dimensional parameter extracted from the above homotopy process, which describes the main coupling direction, bottleneck location and nonlinear characteristics of the influence of different node injections on voltage, current and reactive power distribution.

[0074] Weighted least squares evaluation is an evaluation method that sets weights based on measurement confidence and scenario priority to jointly fit and minimize the error of targets such as voltage deviation, line loss, and harmonics.

[0075] The target evaluation vector is a comprehensive quantitative result generated by weighted evaluation, which gives the score, trade-off coefficient, sensitivity and constraint activity of each target, and is used as a direct input for subsequent optimization.

[0076] Among them, differentiable mixed integer programming is an optimization modeling and solution framework that simultaneously includes continuous variables and discrete decisions and is differentiable for the solution process, making it easy to link with learning modules end-to-end.

[0077] Among them, the dual primal collaborative solution is a solution strategy that collaboratively updates the dual multipliers and the primal decision variables in the same optimization loop, thereby accelerating convergence and enhancing feasibility through mutual correction.

[0078] Specifically, a causal graph (environment → load measurement, environment → device capacity, load → voltage quality, etc.) is constructed using distribution network environmental characteristics as confounding variables. In this graph, environmental characteristics are treated as confounding factors, load characteristics as processing variables, and power quality indicators as results. The original load characteristics are denoised and standardized. Then, propensity score matching or stratification is performed based on environmental vectors such as temperature, altitude, humidity, and dust, and comparable sample pairs are constructed using Mahalanobis distance constraints. Next, a structural causal model is used to establish interpretable causal relationships, introducing stabilization inverse probability weighting and "double robustness" orthogonalization to further offset residual confounding. After determining the causal effects, the environment is fixed as baseline conditions (e.g., rated temperature, rated altitude, and cleanliness) and counterfactual intervention is implemented. A causal representation network or an adversarial debiasing encoder is used to map the "environment-load" relationship to an environment-invariant representation space, and within this space, the counterfactual values ​​of power factor, power change rate, harmonic content, and three-phase imbalance under the baseline environment are inferred. Further, the hypothesis testing of sample coverage and stable unit processing was completed, sensitivity analysis was carried out for unobserved confounding, and equal probability quantile calibration was performed according to temperature zone and altitude zone; finally, the load counterfactual characteristic information was output, that is, the load characteristic vector under the baseline environment without environmental bias, and the confidence interval and effective sample size weight were given.

[0079] The line characteristic information is preprocessed (including temperature and frequency correction of impedance parameters, interpolation of missing measurements and outliers, and verification of switch and segment status). Based on this, a physical constraint graph model with nodes, branches, and switching elements is constructed. Power flow conservation, voltage and current limits, reactive power capacity boundaries, topology connectivity, and safety distances are written into the rule set of the graph as hard constraints. Then, a homotopic continuous path is formulated to gradually approximate the "actual operating network" from the "simplified radial network". The real branch and switch status are enabled segment by segment with small step sizes. A predictive-correction stepping strategy and a reliable domain control are adopted. At each step, physical feasibility and measurement consistency are verified (including the voltage range of key nodes, branch load rate, reactive power margin, and compatibility of protection settings). If necessary, the step size is rolled back or the local topology is sparsified to avoid unsolvable regions. After the homotopy path converges to the target topology, the graph is embedded and sensitivity is extracted to obtain a set of parameters, including the main coupling direction that can characterize the "impact of node injection on voltage and reactive power distribution", the bottleneck branch identifier that is more sensitive to changes in different operating conditions, the local curvature index that reflects the shape of nonlinear response, and the low-dimensional coordinates and segmented reactive power support weights for downstream optimization. These are collectively referred to as line load coupling manifold parameters, and are accompanied by quality metrics (such as residual level, robustness score and interpretability label).

[0080] Load counterfactual characteristics and line load coupling manifold parameters are integrated into an evaluation dataset according to unified dimensions and time alignment rules. The evaluation objects are clearly defined as key power parameters such as voltage deviation, line loss level, harmonic index, and reactive power switching cost. For each parameter, a weight is assigned based on source credibility, scenario priority, and measurement confidence. Then, data cleaning and robustness processing (including outlier suppression, missing measurement interpolation, consistency auditing, and robust standardization) are performed. Collinearity reduction and feature grouping are applied to highly correlated features. The input combination and regularization strength of the evaluation model are determined by combining the sensitivity and bottleneck identification given by the manifold parameters. Based on this, a weighted evaluation solution is performed, outputting a comprehensive result including the fitted values ​​of each objective, residual level, importance ranking, and sensitivity decomposition, and recording the weight contribution, regularization strength, and stability score. Finally, the above results are compressed into an objective evaluation vector, which provides the scores and trade-off coefficients for three types of objectives: voltage quality, energy efficiency, and switching cost; constraint activity indicators; suggestions for the direction and magnitude of key controlled quantities; and uncertainty intervals and diagnostic labels.

[0081] After mapping the target evaluation vector to planning objectives and weights, the two types of decision quantities are clarified (continuous quantities such as DC bus voltage setting, reactive power setting of each section, inverter voltage regulation and harmonic suppression parameters; discrete quantities such as compensation device switching, tap and operating mode selection). Original physical and operational constraints are then loaded (node ​​voltage and branch current boundaries, reactive and active capacity, equipment temperature rise and derating, switching frequency and minimum dwell time, rate of change and total variation penalty, protection and communication constraints, and output range due to environmental reduction). Subsequently, a differentiable mixed-integer programming model is constructed and solved using a primitive-dual collaborative framework. During the solution process, the dual side adaptively strengthens active constraints and provides shadow prices through differentiable Lagrange updates. The primitive side uses differentiable interior points or trust regions for continuous variables to obtain high-quality search directions, and employs smooth approximation and direct estimation for discrete variables to maintain end-to-end differentiability. Simultaneously, neighborhood search and pruning strategies are superimposed to quickly eliminate infeasible combinations. The entire process uses the solution from the previous cycle as a warm start, selects several representative operating conditions for robustness verification according to the uncertainty of the scenario, and performs feasibility and safety audits after each iteration (including voltage / current out-of-bounds checks, temperature rise and life model verification, switching times and minimum dwell constraints verification). After meeting the convergence and feasibility thresholds, the energy demand adaptation analysis data is output, which includes the optimal continuous setpoint and discrete action sequence, constraint active set and shadow price, scenario robustness score and implementation priority.

[0082] In this embodiment, by first performing counterfactual causal correction on load characteristics based on environmental features, biases caused by external disturbances such as temperature, altitude, and dust can be removed, making the load characterization closer to the actual energy demand. Then, by performing physical constraint graph homotopy continuation processing on impedance and topology, the line-load coupling relationship can be stably extracted and bottleneck nodes and reactive power support capacity can be accurately identified. Subsequently, weighted least squares are used to form the target evaluation vector, so that indicators such as voltage quality, line loss, and harmonics can obtain a more reasonable trade-off under the weighting of credibility and priority. Finally, the optimal settings and action sequences that meet the equipment and safety constraints are obtained by dual-primitive co-solution of differentiable mixed integer programming. Under the same power supply capacity, smaller voltage deviation, lower line loss, more controllable harmonics, fewer switching times, and smoother parameter adjustment are achieved, and the robustness and interpretability to extreme environments and measurement uncertainties are improved.

[0083] In an exemplary embodiment, based on energy demand adaptation analysis data, dynamic energy demand adaptation mapping is performed on the power system to obtain energy demand adaptation mapping instructions, including steps 502 to 506. Wherein:

[0084] Step 502: Based on the energy demand adaptation analysis data, perform multi-scenario energy consumption analysis on the power system to obtain energy consumption analysis data for each scenario;

[0085] Step 504: Based on the energy consumption analysis data of each scenario, perform dynamic energy demand strategy analysis on the power system to obtain strategy analysis data for each energy consumption parameter;

[0086] Step 506: Perform strategy fusion mapping on the strategy analysis data of each energy consumption parameter to obtain energy demand adaptation mapping instructions.

[0087] Among them, the multi-energy consumption scenario analysis divides and labels scenarios according to the dimensions of load type, time period, environmental level and equipment status based on the energy demand adaptation analysis data, and extracts the target priority, constraint list and key controlled quantities of each scenario.

[0088] Dynamic energy demand strategy analysis evaluates the marginal benefits and costs of control parameters for a single scenario under feasible domain and action budget constraints, and generates an executable parameter strategy solution.

[0089] Energy parameter strategy analysis data is a structured result set obtained from strategy analysis, including candidate parameter set, expected indicator improvement amount, execution cost, robustness score and effective conditions, etc.

[0090] Policy fusion mapping involves resolving conflicts and weighting candidate policies from multiple scenarios, then projecting them back to system constraints to generate the final energy demand adaptation mapping instruction.

[0091] Specifically, the target weights, active constraint sets, and uncertainty indicators in the adaptation analysis data are clustered by time and geographical units. A scenario labeling system is established by combining load type, operating period (peak / valley / impact), environmental level (low temperature / high dust / high altitude), and equipment status (capacity margin, temperature rise, alarms). Scenario matching and conflict verification are performed on each record (e.g., conflict between voltage quality priority and switching cost), generating a ternary description of "scenario-constraint-target." Within each scenario, the sensitivity, feasible domain boundary, and risk score of key controlled quantities are summarized to form energy consumption analysis data for each scenario (including scenario ID, target ranking, constraint list, recommended controlled variables, and risk level).

[0092] For each scenario's corresponding energy consumption analysis data, the target ranking and constraint list are retrieved. Voltage quality-related variables (such as DC bus voltage setting, segmented SVG reactive power command, and inverter voltage regulation / harmonic suppression parameters) are prioritized and action budgets (number of switching operations, rate of change, minimum dwell time) are applied. Within the feasible region, parameter scanning or rapid heuristic optimization is performed to evaluate the marginal benefits to voltage deviation, line loss, harmonics, and execution costs. If necessary, extreme samples are introduced for robustness verification and rollback strategy setting. Energy consumption parameter strategy analysis data is output, including candidate parameter set, expected indicator improvement, execution cost, robustness score, and effective conditions.

[0093] Candidate strategies for all scenarios are used to construct a conflict graph based on their activation conditions, time priorities, and mutual exclusion relationships. Strategies with resource contention or action conflicts are first screened for compatibility, and then weighted voting and consistency verification are performed based on target weights and robustness scores. Final constraint projections (voltage / current boundaries, capacity limits, and action throttling) are applied to the strategy combinations that pass the verification, and standardized energy demand adaptation mapping instruction messages are generated, specifying the set values ​​(such as Udc*, Qi* for each segment, inverter control parameters, and environmental action thresholds), activation windows, rollback conditions, and monitoring points for each instruction.

[0094] In this embodiment, by first decomposing the adaptation analysis data into multiple scenarios and extracting their respective objectives and constraints, then generating targeted parameter strategies within each scenario, and finally forming unified instructions through cross-scenario conflict resolution and weighted selection, the targeting and interpretability of the strategy can be significantly improved. This further reduces voltage deviation and line loss, makes harmonic suppression more effective, reduces invalid switching and parameter jitter, shortens the response time from identification to decision, takes into account the robustness of extreme environments and measurement uncertainties, and clarifies the effective conditions and priorities with standardized instructions, thereby bringing better energy efficiency indicators, equipment lifespan and power supply reliability.

[0095] In an exemplary embodiment, based on energy consumption analysis data for each scenario, dynamic energy demand strategy analysis is performed on the power system to obtain strategy analysis data for each energy consumption parameter, including steps 602 to 608. Wherein:

[0096] Step 602: Perform counterfactual perturbation analysis on the energy consumption analysis data for each scenario to obtain a scenario counterfactual sample set;

[0097] Step 604: Based on the scenario counterfactual sample set, perform extreme value optimization within the distributed feasible region for each candidate energy consumption parameter of the power system to obtain the candidate energy consumption parameter set;

[0098] Step 606: Perform a weighted least squares consistency check on the candidate energy consumption parameter set to obtain a compliant candidate parameter set;

[0099] Step 608: Based on the set of compliant candidate parameters and the set of scenario counterfactual samples, perform differentiable mixed integer policy distillation mapping on the policy space of the power system to obtain policy analysis data for each energy consumption parameter.

[0100] Counterfactual perturbation analysis involves generating representative perturbations based on a hypothetical "what if / what if" scenario and replaying the system response to evaluate the robustness of a strategy under different external conditions.

[0101] Among them, the scenario counterfactual sample set is a sample set obtained from counterfactual disturbance analysis, which includes each disturbance setting and its corresponding voltage, line loss, harmonics, cost and other responses and feasibility labels.

[0102] Among them, the candidate energy consumption parameters are the specific values ​​of parameters and actions that can be actually issued under given scenarios and constraints (such as Udc settings, reactive power commands of each segment of SVG, inverter control parameters, switching and gear selection).

[0103] Among them, the extreme value search within the feasible region is a method that searches for the optimal combination of parameters that makes the target perform best and remains feasible on most samples, taking into account the perturbation distribution and feasible constraints.

[0104] Among them, the candidate energy parameter set is a summary of multiple high-potential parameters selected by extreme value optimization within the feasible distribution domain, as well as their cross-sample performance statistics and robustness information.

[0105] The weighted least squares consistency check is a process that evaluates the weighted residuals and stability of candidate parameters on each target according to sample weights and confidence levels to determine their consistency with the scene target.

[0106] Among them, the compliant candidate parameter set is a subset of candidate parameters that meet the performance and feasibility thresholds after consistency verification and are marked with applicable conditions and active constraints.

[0107] The strategy space is a search domain consisting of all feasible parameters and discrete actions and their combinations, and is constrained by physical, equipment, and operational constraints.

[0108] Among them, the differentiable mixed integer policy distillation mapping is a process of compressing the excellent decision rules of compliant candidate parameters into a policy model that contains both continuous settings and discrete actions, and whose solution process is differentiable, so as to output executable instructions online quickly.

[0109] Specifically, using energy consumption analysis data from various scenarios as seed records, a "list of exogenous disturbances" is compiled (environmental levels such as low temperature / high dust / high altitude, measurement noise and offset, sudden load jumps and continuous fluctuations, equipment status changes such as reactive power capacity decay or switching lag), and scenario baseline configurations and constraint boundaries are established. Then, reasonable amplitude ranges, rates of change, and probabilities of occurrence are set for each type of disturbance. Orthogonal or Latin hypercube sampling is used to generate representative disturbance combinations, and they are stratified into mild / moderate / severe to ensure coverage and observability of extreme conditions. For each disturbance combination, it is replayed in layers between the fast power flow core and the electromagnetic transient simulator (first using a simplified model to screen feasible and obviously failed samples, and then using a high-precision model to verify the "critical samples"). Response quantities such as node voltage out-of-bounds conditions, line loss and harmonic indicators, SVG reactive power output and inverter control saturation, action costs and temperature rise risks are uniformly recorded, and feasibility indicators and triggered constraints are marked. Automatic reruns and adaptive step sizes are performed for abnormal or unstable simulations. After removing duplicate and redundant samples, stratified reweighting is carried out to make the sample distribution fit the actual probability of occurrence of the scenario. At the same time, metadata (disturbance source, intensity level, time location, credibility score, recommendation rollback strategy) is generated for each sample. Finally, a counterfactual sample set of the scenario is produced, covering representative settings from perturbations to extremes, with weights, risk labels and reproduction parameters.

[0110] Based on a scenario-based counterfactual sample set, boundaries, resolution, and rate-of-change limits are set for each type of adjustable parameter (such as the upper and lower limits and steps of DC bus voltage setting, the minimum scheduling granularity of SVG reactive power commands for each segment, the safety domain of inverter voltage regulation and harmonic suppression parameters, the number of switching operations, and the minimum dwell time), and baseline configuration and historical good values ​​are used as a warm start. A rapid evaluation link is then constructed, which involves batch replaying each candidate parameter combination on the counterfactual sample set according to sample weights. First, a lightweight approximation model is used for coarse screening, and then the combinations that enter the top ranks are verified using a high-precision power flow and harmonic evaluation kernel. Simultaneously, constraint checks are performed (voltage and current boundaries, capacity and temperature rise, throttling operation, communication and protection coordination). Combinations that violate constraints are directly discarded or constraint projection corrections are performed. The search strategy adopts a hybrid approach of heuristic and degenerate bar parallelism, combining local gradient guidance, neighborhood perturbation, random restart, and population evolution. It prioritizes solutions that remain feasible and have stable target improvement on most perturbation samples, and penalizes solutions that are highly sensitive to environmental or measurement uncertainties. To avoid overfitting to a single extreme sample, robustness and consistency scores are introduced as ranking criteria, and early stopping and diversity maintenance mechanisms are set up to ensure that candidate solutions achieve a balance between performance, feasibility, and implementable cost. The final output is a set of candidate energy parameters, with each candidate solution accompanied by cross-sample performance statistics, a list of active constraints, robustness and consistency scores, recommended effective conditions, and rollback suggestions.

[0111] The response results of the candidate energy parameters set on all counterfactual samples are summarized into a unified verification data table. Weights are assigned to each sample based on the credibility of the sample source, scenario priority, measurement confidence, and extreme degree. Missing measurements and abnormal records are imputed, denoised, and audited for consistency. Then, evaluation items and tolerances (such as voltage deviation, line loss, harmonics, switching costs, and temperature rise risk) and alarm thresholds are set for the target side. A hierarchical weighting strategy is adopted to highlight high-priority targets and key nodes. Robust handling is introduced to suppress the influence of a small number of outliers, and smoothing constraints are applied to temporally adjacent samples to avoid unreasonable jumps in the strategy's time series. Based on this, weighted residuals and stability indices are calculated for each candidate solution, and the output is ranked and labeled. The sources of residuals are interpreted and decomposed to identify whether the main mismatch is caused by model error, measurement uncertainty, or the parameters themselves. Candidate solutions that do not meet the tolerance or are continuously unstable in specific sub-scenarios are eliminated or downgraded. For candidate solutions that are critical but can be fine-tuned to enter the qualified range, suggested correction amounts are given and re-verified. Finally, a set of compliant candidate parameters is formed. Each candidate solution that passes the verification is accompanied by a credibility score, a list of active constraints, applicable conditions and prohibited conditions, suggested observation points and re-examination cycle.

[0112] Using the set of compliant candidate parameters and their performance on a counterfactual sample set of scenarios as teacher signals, a student policy model is constructed that simultaneously covers continuous settings (such as DC bus voltage, segmented reactive power settings, inverter voltage regulation and harmonic suppression parameters) and discrete actions (such as compensation device switching, gear selection, and operating mode selection). The unified input is scenario features, risk labels, and constraint summaries. Then, batch distillation training is performed, that is, guided by the improvement of the teacher solution's indicators, robustness score, and consistency score, the student model's fitting priority for high-value decisions is optimized. Hard constraints such as voltage and current limits, capacity and temperature rise limits, action throttling and minimum dwell time, and communication and protection coordination are embedded into the forward and backward processes through a constraint projection layer to ensure that the output is naturally executable. To reduce overfitting and fragility, sample reweighting and perturbation enhancement strategies are introduced to force the student policy to maintain stable performance on mild, moderate, and severe perturbation subsets, and smoothing and monotonicity priors are applied to sensitive features. In the later stages of training, confidence calibration and threshold tuning are performed to ensure that the policy output simultaneously provides the setpoint, action selection, expected metric benefit, confidence interval, trigger condition, and fallback path. Finally, a safety and feasibility audit is completed on an independent validation set (including critical node out-of-bounds checks, temperature rise and lifetime model verification, and switching frequency and minimum dwell time constraint verification). The audited policies are then solidified in a structured form as energy consumption parameter policy analysis data, which includes recommended settings and action combinations for each scenario, applicable boundaries, monitoring points, and re-inspection cycles.

[0113] In this embodiment, counterfactual perturbation analysis is first performed on the scenario data to cover uncertainties in the environment, measurement, and load. Then, extreme value optimization is performed within the distributed feasible domain to screen out parameter combinations that remain feasible and have stable returns under most perturbations. Subsequently, weighted least squares consistency check is used to eliminate schemes that are sensitive to a small number of abnormal samples or have imbalances across indicators. Finally, compliant candidates are distilled and compressed into a directly deployable strategy mapping through differentiable mixed integer strategy. This significantly improves the comprehensive optimization effect of indicators such as voltage deviation and line loss without increasing hardware costs, enhances robustness to extreme operating conditions and noise, reduces invalid switching and parameter jitter, reduces online solution latency, and improves the interpretability and maintainability of the strategy.

[0114] In an exemplary embodiment, strategy fusion mapping is performed on the strategy analysis data of each energy consumption parameter to obtain energy demand adaptation mapping instructions, including steps 702 to 708. Wherein:

[0115] Step 702: Perform dual embedding representation processing on the strategy analysis data of each energy consumption parameter to obtain a set of strategy dual representations;

[0116] Step 704: Based on the strategy dual representation set, perform zero-knowledge game analysis on the conflicting strategies in the strategy analysis data of each energy parameter to obtain a subset of mutually compatible strategies.

[0117] Step 706: Perform weighted voting fusion on the mutually compatible strategy subset to obtain preliminary fusion strategy parameters;

[0118] Step 708: Perform neural satisfiability-driven strategy synthesis on the preliminary fusion strategy parameters to obtain energy demand adaptation mapping instructions.

[0119] Among them, dual embedding representation processing is the process of encoding and aligning the original spatial information of "goal-cost-scenario" and the dual information of "constraint-tension-sensitivity" of each strategy into a unified vector.

[0120] Among them, the policy dual representation set is the complete policy vector library obtained after dual embedding representation, with interpretable labels and retrieval indexes for subsequent filtering and comparison.

[0121] Among them, conflict strategies are those that are mutually exclusive or mutually restrictive on the same time window or the same resource, making it difficult to simultaneously satisfy the target benefits and constraints.

[0122] Zero-knowledge game analysis is an analytical method that evaluates the resource consumption and benefit claims between strategies through commitment and verification interactions without revealing the details of the strategies, and determines whether they can coexist.

[0123] The mutually compatible strategy subset is a set of strategies that, after zero-knowledge game analysis, can coexist in terms of resources and constraints, and provide necessary concessions and effective window descriptions.

[0124] Weighted voting fusion is a method that assigns vote weights to strategies based on target weights, robustness and credibility scores, execution costs, etc., and combines continuous weighted aggregation with majority decision-making for discrete actions to form a composite result.

[0125] The initial fusion strategy parameters are the first version of parameters and action combinations obtained by weighted voting fusion, including source weight decomposition, expected returns, and applicable conditions.

[0126] Among them, neural satisfiability-driven strategy synthesis is a process of using a neural scorer to guide symbolic satisfiability solving, projecting preliminary parameters onto the hard-constrained feasible domain and repairing local conflicts to generate the final instruction.

[0127] Specifically, for each energy consumption parameter strategy analysis data, elements are sorted and standardized, dimensions are unified and missing items are filled, and original elements such as target benefits, execution costs, robustness scores, active constraint sets, key controlled quantities, and applicable scenarios are extracted. Dual-side information is derived based on physical and operational rules, including constraint tension, shadow price approximation, sensitivity to key controlled quantities, bottleneck node identification, forced concession costs, and potential conflict resource occupancy rates, and outlier suppression and interval pruning are performed. Based on this, two encoders are constructed: the original spatial encoder embeds "target—cost—scenario—controlled quantity," and the dual spatial encoder embeds "active constraints—tension—sensitivity—shadow price approximation—bottleneck information." Alignment training and mutual information regularization are used to improve the consistency and distinguishability of the two representations, while class balancing, feature desensitization, and privacy labels are added to avoid leaking sensitive parameters. Then, dimensionality reduction and visualization separability checks, clustering and class purity assessments, stability resampling, and cross-period hold-out validation are performed to remove unstable representations and gently smooth high-variance dimensions. Finally, the two embeddings are fused into a unified vector according to learnable weights, generating interpretable labels and fast retrieval indexes for each policy, including the dominant objective, main constraints, bottleneck location, risk level and credibility, forming a policy dual representation set.

[0128] A conflict graph is constructed based on the strategy duality representation set, clarifying the potential mutual exclusion relationships of each strategy in terms of time window, equipment resources (DC bus, SVG segments, inverter channels), operation and maintenance quotas (switching times, minimum dwell time), and safety boundaries (voltage, current, temperature rise, protection settings). A commitment summary (containing only the target benefit range, active constraint labels, resource occupancy fingerprints, and credibility) is generated for each strategy without revealing details. Following a challenge process of "initiation-response-adjudication," candidate strategies verify each other's feasibility and benefit claims using commitment summaries. Key claims are verified through random sampling and external auditors, recording the minimum concession costs, alternative time windows, degradeable configurations, and fallback paths for each party under resource conflicts, constraint triggering, and benefit constraints. Verified interaction records are aggregated, and conflict edges are pruned or downweighted using rule priorities (safety > power quality > energy efficiency > cost), scenario levels, and credibility scores. Sets with chained conflicts are decomposed and layered (high-risk conflicts are eliminated first, then medium-risk conflicts are coordinated, and finally low-impact differences are retained). Upon completion, a subset of the reciprocal policies will be output, with each retention policy accompanied by a reciprocal specification (including the resource allocation allowed for coexistence, effective time window, necessary concessions and disabling conditions), conflict resolution basis, audit fingerprint, and review cycle.

[0129] The subset of mutually compatible strategies undergoes unified preprocessing (uniformity of dimensions, time alignment, and repair of missing and anomalies), and voting weights are assigned to each strategy based on global and scenario-level target priorities, robustness scores, reliability scores, execution costs, and historical stability. Then, fusion is carried out by grouping according to resource and time dimensions. During fusion, continuous settings (DC bus voltage, segmented reactive power settings, inverter voltage regulation and harmonic suppression parameters) are weighted and aggregated, with rate-of-change limits, smoothing filters, and minimum dwell constraints applied to suppress jitter. Discrete actions (compensation device switching, gear selection, and mode selection) are decided by weighted majority vote, with tie-breaking rules set (safety first, power quality second, energy efficiency third, cost last). The preliminary aggregation results are subjected to hierarchical constraint projection, and voltage and current boundaries, capacity and temperature rise limits, action throttling and protection coordination, and communication and execution path reachability are verified item by item. If conflicts occur, local items are rolled back or replaced according to the principle of minimizing weight and impact. Simultaneously, historical consistency audits and cost-saving strategy checks are conducted to avoid frequent back-and-forth with recently issued instructions. The final output is a preliminary fusion strategy parameter, which includes various settings and action selections, source weight decomposition, expected indicator returns and uncertainty ranges, applicable time windows and rollback conditions.

[0130] The initial fusion strategy parameters are mapped to variables and constraints encoding a satisfiability problem. Constraints cover voltage and current boundaries, reactive and active capacity, temperature rise and derating conditions, switching frequency and minimum dwell time, protection and communication coordination, and scenario activation windows. A collaborative process consisting of a neural scorer and a symbolic solution kernel is then constructed. The neural scorer assigns values ​​to variables and scores subsets of constraints based on historical successful and conflicting samples, prioritizing combinations that are more likely to be satisfied and offer higher returns. The symbolic solution kernel performs consistency checks on candidate solutions and returns conflict explanations and remediation clues. For unsatisfied localities, guided conflict learning and candidate replacement are performed. When necessary, minor concessions or time window rearrangements are made to individual settings according to a hierarchical strategy of "safety priority—power quality—energy efficiency—cost." After each update, a feasibility audit of power physics and operation and maintenance rules is invoked (including critical node boundary violations, capacity and temperature rise verification, and action throttling and protection coordination verification). Once both the satisfaction score and the benefit score reach the threshold and pass independent verification and regression testing, a structured energy demand adaptation mapping instruction is generated, which clarifies the DC side target voltage, segmented reactive power settings, inverter voltage regulation and harmonic suppression parameters, environmental action thresholds and their effective periods, priorities, monitoring points, triggering and rollback conditions, and also includes a constraint active list, key evidence and audit fingerprint.

[0131] In this embodiment, by using dual embedding representation of strategy analysis data, key information such as "objective-cost-constraint-sensitivity" can be compressed into a searchable and interpretable vector representation to improve discrimination and recall. On this basis, zero-knowledge game analysis is used to identify and resolve conflicting strategies across resources and time windows, protecting sensitive details while reducing performance loss caused by mutual exclusion. Subsequently, weighted voting fusion is used to combine robustness and cost under the premise of prioritizing safety and power quality, resulting in preliminary parameters with low jitter and high feasibility. Finally, through neural satisfiability-driven strategy synthesis, the fusion result is projected into the hard-constraint feasible domain and local conflicts are automatically repaired, thereby achieving a faster instruction generation, lower risk of out-of-bounds movement, more robustness to extreme conditions, and stronger traceability of energy demand adaptation mapping output.

[0132] In an exemplary embodiment, the power system's operating strategy is optimized and adjusted based on distribution network operation status data to obtain dynamic control data for distribution network energy consumption, including steps 802 to 808. Wherein:

[0133] Step 802: Perform time-series comparison learning representation processing on the distribution network operation status data to obtain the operation status vector;

[0134] Step 804: Based on the operating status vector, perform topology sensitivity sparsification analysis on the adjustable points in the power system to obtain the set of adjustable points;

[0135] Step 806: Perform weighted least squares correction on the voltage and reactive power measurement data in the adjustable point set to obtain the initial parameter adjustment amount;

[0136] Step 808: Based on the initial parameter adjustment amount and the operating status vector, perform differentiable robust model predictive control solution for the power system to obtain dynamic control data for distribution network energy consumption.

[0137] Among them, temporal contrastive learning representation processing is the process of obtaining low-dimensional feature representations that can be used for downstream decision-making by enhancing and contrasting the multi-channel runtime sequence to distinguish different operating conditions.

[0138] Among them, the operational status vector is a compact feature vector obtained from time-series representation, which comprehensively reflects the current steady state, fluctuations and risk clues of the distribution network.

[0139] Among them, adjustable points are network nodes that, under the constraints of the current topology and equipment, have a significant impact on the target (such as voltage quality and line loss) and have practical controllable means.

[0140] Among them, topology sensitivity sparsification analysis is a method to evaluate the influence of each node on the controlled variable and perform sparsity screening in combination with physical and operational constraints to retain the nodes with the most controllable value.

[0141] The adjustable point set is a list of executable controllable nodes that have been screened for sensitivity and feasibility and arranged by priority.

[0142] Among them, reactive power measurement data consists of reactive power and related status data of each node / equipment collected by on-site measurement or metering devices.

[0143] Weighted least squares correction is a method that assigns weights based on measurement confidence and priority to perform consistency correction on measurements such as voltage and reactive power to obtain more reliable estimates.

[0144] Among them, the initial parameter adjustment amount is the first round of adjustment suggestions for each set value or control quantity after calibration, which is used as the controller's warm start.

[0145] Among them, differentiable robust model predictive control solution is a control solution process that, under the premise of considering uncertainties and hard constraints, makes the rolling optimization problem differentiable and obtains the optimal settings and actions in real time.

[0146] Specifically, the real-time and near-historical time-series data of the distribution network operation status, including voltage, current, reactive power, power factor, harmonics, line loss, temperature rise, actuator status, and alarms, are sliced ​​according to fixed windows and step sizes. After missing data repair, denoising, and standardization, multi-view samples (original sequence, frequency domain summary, event markers, environmental labels, etc.) are constructed. Positive and negative sample pairs are generated using enhancement and negative enhancement rules, and a contrastive representation network is trained to aggregate similar operating conditions in the feature space and separate dissimilar operating conditions. Then, forward inference is performed on the latest data to obtain a low-dimensional operating status vector that takes into account steady-state, fluctuation, and risk clues, while outputting the representation confidence and anomaly score.

[0147] The operational status vector, along with the current network topology, equipment capabilities, and constraints, is input into the sensitivity analysis model to evaluate the impact of each node's voltage on controllable variables (such as segmented reactive power, DC-side voltage settings, and inverter parameters). Feasibility and safety boundary screening (capacity, temperature rise, switching budget, minimum dwell time, etc.) are then overlaid. Through sparsification and grouping constraints, only nodes that significantly contribute to the improvement of the target and have the conditions for execution are retained, forming a priority-ranked set of adjustable points. A description of the recommended control objects and impact range for each node is also provided.

[0148] Consistency auditing and anomaly suppression are performed on the measurement data of the adjustable point set. Based on the measurement confidence, sensor calibration information and historical stability, weights are assigned to construct a correction problem including voltage deviation, reactive power output and power factor. Measurement points with obvious distortion are eliminated or reduced in weight. In combination with the rated and operating boundaries of the equipment, reasonable rate of change constraints are applied to obtain a set of initial parameter adjustment quantities that meet physical consistency and can converge quickly. At the same time, active constraints and suggested observation points are recorded to provide a high-quality initial solution for predictive control.

[0149] The initial parameter adjustments are used as a warm start. Uncertainty and priority information extracted from the operating status vector are injected into the prediction model and cost weights to construct a robust model predictive control problem that includes constraints such as voltage and current boundaries, reactive and active capacity, temperature rise and derating, switching frequency and minimum dwell time, and protection and communication coordination. The solver is made differentiable and configured in the rolling time domain for real-time updates. The optimal settings and actions for the current cycle (DC bus voltage target, segmented reactive power distribution, inverter voltage regulation and harmonic suppression parameters, environmental action thresholds, etc.) are obtained along the time axis, with the addition of rate of change limits, dead zones, and backoff strategies, forming dynamic control data for distribution network energy consumption.

[0150] In this embodiment, the operational status vector is first extracted using time-series comparative learning to improve the accuracy and timeliness of operational condition judgment. Then, topology sensitivity sparsification is performed based on this vector to retain only the most controllable adjustable points. Next, weighted least squares correction is used to stabilize key measurements and provide highly reliable first-round parameter tuning. These reliable priors are then injected into the real-time solution of the differentiable robust model predictive control. This achieves faster decision convergence, smaller voltage deviation and line loss, lower harmonics and switching frequency under the same hardware conditions, and significantly enhances robustness and interpretability to environmental disturbances and measurement uncertainties, while reducing online computational burden and strategy jitter.

[0151] In an exemplary embodiment, based on the initial parameter adjustment and the operating status vector, a differentiable robust model predictive control is performed on the power system to obtain dynamic control data for distribution network energy consumption, including steps 902 to 910. Wherein:

[0152] Step 902: Based on the operating situation vector and the initial parameter adjustment, the disturbance set in the prediction model of the differentiable robust model predictive control is modeled using the Wasserstein sphere distribution to obtain the robust uncertainty set parameters.

[0153] Step 904: Perform soft-constrained weighted least squares tuning on the prediction cost function of the differentiable robust model predictive control solution to obtain the soft-constraint weights and the initial Lagrange multipliers.

[0154] Step 906: Based on the robust uncertainty set parameters and soft constraint weights, construct the model predictive control information of the prediction model in a quadratic programming form to obtain a quadratic programming controller.

[0155] Step 908: Perform implicit layering differentiability parameterization on the quadratic programming controller to obtain a differentiable quadratic programming controller.

[0156] Step 910: Based on the differentiable quadratic programming controller, perform dual disturbance alignment solution within the time domain rolling window to obtain the dynamic control data of power distribution network energy consumption.

[0157] Among them, the prediction model refers to the dynamic model used to predict the changes in the distribution network status with control variables and external conditions, which includes elements such as controlled / measuring channels, reference trajectories, and terminal feasible regions.

[0158] The disturbance set refers to a unified set of uncertain sources such as load, measurement, model error, and environmental reduction, which is used to inject disturbance scenarios into prediction and constraint verification.

[0159] Among them, Wasserstein sphere distribution modeling is a robust modeling method that uses the sample distribution as the center and constructs a neighborhood with a given radius to characterize the distribution offset range.

[0160] Among them, the robust uncertainty set parameter is a configuration parameter that describes the shape, radius, source weight and sampling strategy of the disturbance set, and is used to guide robust prediction and constraint tightening.

[0161] The prediction cost function is a comprehensive loss function that measures objectives such as voltage quality, line loss, harmonics, and execution cost within the prediction time domain.

[0162] Among them, soft-constrained weighted least squares tuning is a method of tuning the strength of soft constraints by assigning weights to the out-of-bounds penalty term and the objective term and minimizing the weighted error.

[0163] Among them, soft constraint weights are penalty weights and tolerance parameters imposed on various out-of-bounds and target items, used to balance safety, quality and cost.

[0164] The initial Lagrange multipliers are the initial values ​​of the dual variables used to characterize the activity of the constraints, which help accelerate the convergence of subsequent optimizations.

[0165] Among them, model predictive control information is a unified configuration list that integrates uncertainty set, target weights, constraints and time domain settings, which is used for controller construction and solution.

[0166] Among them, the construction of the quadratic programming form is the process of expressing the predictive control objective and linear constraints as a standard optimization form of a quadratic objective function and linear constraints.

[0167] Among them, the quadratic programming controller is an MPC solver configuration implemented according to the quadratic programming form, which can output the optimal control quantity at each time step.

[0168] Among them, the differentiable parameterization of implicit layering is a technique that embeds the solution process into a differentiable implicit layer and establishes smooth dependencies on key hyperparameters.

[0169] Among them, the differentiable quadratic programming controller is a QP controller that has been made differentiable, which can be jointly adjusted end-to-end with the upstream model and stabilize backpropagation.

[0170] Among them, the time-domain rolling window is an operating mechanism that repeatedly predicts, optimizes, and executes within a finite future time domain and moves forward over time.

[0171] Among them, the dual perturbation alignment solution is a solution strategy that updates the original variables and dual multipliers simultaneously during optimization and checks the high-risk perturbations first, so that the solution converges in a coordinated manner on both the feasible and robust sides.

[0172] Specifically, uncertainty clues such as load fluctuation amplitude, rate of change, measurement noise level, and environmental level are extracted from the operational status vector. Combined with the executable range and historical deviations reflected by the initial parameter adjustments, disturbance types (load disturbance, measurement disturbance, model error, and environmental reduction) and their scales are defined. Based on this, the radius, shape, and weight sources of the uncertainty set are set using a Wasserstein sphere robust modeling approach, forming robust uncertainty set parameters that can be implemented (including disturbance sources, intensity upper limits, confidence stratification, and sampling strategies). An interface-based configuration for constraint verification is then output.

[0173] Based on current operational goals and maintenance strategies, evaluation items such as voltage deviation, line loss, harmonics, reactive power switching costs, rate of change penalties, and minimum dwell constraints are determined. Weighting coefficients are set based on measurement confidence and scenario priority. Short-term unavoidable out-of-bounds errors corresponding to the prediction cost function solved by the differentiable robust model predictive control are penalized and tolerances are provided in the form of soft constraints. Then, the soft constraint weights are tuned through historical data playback and small-scale online trials, and the Lagrange multipliers are initialized as priors for constraint activity levels.

[0174] Based on the robust uncertainty set parameters and soft constraint weights, the state prediction, control variables, target weights, and constraint list of the prediction model are compiled into a standard quadratic objective and linear constraint description. Voltage and current boundaries, equipment capacity and temperature rise limits, action throttling and minimum dwell time, protection and communication coordination, and robustness check terms caused by uncertainty sets are explicitly included. A controller description file for quadratic programming solution is generated as a quadratic programming controller. This description also includes the time-domain rolling window length, sampling step size, warm-start position, and backoff rules.

[0175] Without altering the external interface of the quadratic programming controller, its internal solution process is embedded as an implicit layer, exposing its differentiability dependencies on high-level hyperparameters such as objective weights, constraint thresholds, and the radius of the uncertainty set. Numerical stabilization, scaling, and damping strategies ensure the solution is differentiable and insensitive to initial conditions. Furthermore, safety hooks (fast bounds detection, backtracking to the previous feasible solution, and freezing the rate of change of sensitive variables) are added to ensure robustness and interpretability during online operation, ultimately resulting in a differentiable quadratic programming controller.

[0176] The process proceeds sequentially across the rolling time domain. In each control step of the differentiable quadratic programming controller, the latest measurements and predictions are stitched together to form the current window data. Based on the differentiable quadratic programming controller, the original variables and dual multipliers are updated synchronously, and disturbance sampling is aligned (prioritizing coverage near active constraints and high-risk scenarios). Hard conditions such as voltage and current out-of-bounds, capacity and temperature rise, and coordination of action throttling and protection are checked in real time. When the convergence and feasibility criteria meet the thresholds, the dynamic control data of distribution network energy consumption for this cycle is output, including the DC bus voltage target, segmented reactive power distribution, inverter voltage regulation and harmonic suppression parameters, and necessary environmental action thresholds and throttling settings.

[0177] In this embodiment, by characterizing the disturbance using the Wasserstein sphere method and tuning the soft constraint weights and dual initial values, the prediction model, after being constructed as a quadratic programming model, can balance safety margin and objective trade-offs. This allows for priority verification of high-risk disturbances and primary-dual co-convergence in differentiable and rolling solutions, resulting in faster computational convergence and lower out-of-bounds rates. This significantly reduces voltage deviation and line loss, suppresses harmonics and unnecessary switching, reduces parameter jitter, and improves robustness to environmental and measurement uncertainties. At the same time, it retains interpretable outputs on constraint activity and shadow prices, facilitating verification and parameter tuning by the operation and maintenance side.

[0178] In a specific embodiment, (1) the rectifier host module, as the core of the front-end power conversion, is responsible for converting the AC power of the low-voltage distribution network into DC power, and at the same time realizes line impedance monitoring and dynamic voltage regulation. It consists of a power conversion unit, a control and regulation unit, an overcurrent protection unit and a heat dissipation unit:

[0179] ① Power conversion unit: 650V / 400A enhancement-type IGBTs are selected, adopting a three-phase two-level topology. The junction temperature range of the IGBT chip is -40℃~150℃, and the conduction loss is reduced by 20% compared with traditional IGBTs, adapting to a load fluctuation range of 100-180kW. A 1000V / 500A fast recovery diode is connected in series in the topology circuit to prevent reverse voltage from breaking down the IGBT, and a 0.1μF / 1000V high-frequency filter capacitor is connected in parallel to suppress voltage spikes.

[0180] ② Control and Adjustment Unit: Based on the STM32F407 microcontroller, it integrates a real-time line impedance monitoring circuit, incorporates adaptive PI control + fuzzy compensation algorithm, with an algorithm iteration cycle ≤1ms and dynamic response time ≤8ms; by real-time acquisition of line current and voltage signals, it calculates the line impedance value and adjusts the IGBT conduction angle according to impedance changes, achieving a voltage retention rate ≥91.2% at 3km;

[0181] ③ Overcurrent protection unit: It adopts a dual protection mechanism of hardware + software. At the hardware level, it integrates a 600A fast fuse with a response time of ≤10μs; at the software level, it incorporates overcurrent threshold judgment logic and has overvoltage, undervoltage and overtemperature protection functions.

[0182] ④ Heat dissipation unit: It adopts a composite heat dissipation structure of "aluminum heat sink + axial fan + heat pipe". The heat sink area is 0.5m² and the thickness is 5mm. The surface is sprayed with a heat dissipation coating with a thermal conductivity of 0.8W / (m・K); the axial fan speed is 2000r / min and the air volume is 120m³ / h. It has a temperature adaptive speed adjustment function; the heat pipe adopts a copper gravity heat pipe with a diameter of 8mm and a thermal conductivity of ≥150W, ensuring that the IGBT operating temperature is ≤85℃ in an environment of -30℃~45℃.

[0183] (2) Inverter slave module, deployed on the load side, is responsible for converting DC power into AC power that meets national standards, ensuring the power quality of the end load. It consists of a power inverter unit, a control and voltage regulation unit, a harmonic suppression unit, and a low temperature adaptation unit.

[0184] ① Power inverter unit: It adopts 900V / 180A wide bandgap SiC enhancement IGBT, with full-bridge inverter topology. The SiC IGBT breakdown voltage is ≥1200V, the on-state voltage drop is stable at 1.8V at -30℃, and the inverter efficiency is ≥96.1%. The output of the topology circuit is connected in parallel with an LC filter circuit to suppress high-frequency harmonics, and a 200A AC contactor is connected in series to realize load switching control.

[0185] ② Control and voltage regulation unit: The TI TMS320F28335 DSP is used as the control core. It adopts space vector pulse width modulation (SVPWM) + third harmonic injection technology. The SVPWM carrier frequency is 10kHz and the third harmonic injection amount is 15%. Combined with voltage-current dual closed-loop PI control, the voltage loop proportional coefficient Kp=3.2 and integral coefficient Ki=1.5, the current loop proportional coefficient Kp=0.8 and integral coefficient Ki=0.3, the terminal voltage adjustment accuracy is ≤±1.8%, and the voltage recovery time is ≤12ms.

[0186] ③ Harmonic Suppression Unit: Integrates active power filter (APF) function, adopts harmonic detection algorithm based on instantaneous reactive power theory, can compensate for 2nd-31st harmonics, harmonic current compensation accuracy ≤ ±5%, total harmonic distortion (THD) ≤ 2.9%;

[0187] ④ Low-temperature adaptation unit: A 100W silicone heating element is deployed near the IGBT module and equipped with a temperature sensor (measurement range -50℃~100℃, accuracy ±0.5℃). When the ambient temperature is ≤-10℃, the heating element will automatically start to preheat the IGBT module to above -5℃, ensuring a 100% success rate for low-temperature startup.

[0188] (3) Communication module, which realizes data interaction and remote monitoring between modules, has the characteristics of anti-interference, low power consumption and wide coverage, and consists of physical layer transmission unit, protocol and control unit, signal enhancement unit and low power control unit:

[0189] ① Physical layer transmission unit: Utilizes an NB-IoT chip, supports 800 / 900MHz frequency bands, maximum transmit power 23dBm, and receive sensitivity ≤-129dBm; Equipped with an omnidirectional fiberglass antenna, 30cm in length, fixed at a height of 1.2m on top of the device, using an IP68 waterproof connector for connection, reducing signal attenuation by 40% when penetrating shrubs and low buildings, and achieving a signal reception strength ≥-95dBm at 3km;

[0190] ② Protocol and Control Unit: Utilizing an STM32L431 microcontroller as the control core, it incorporates an adaptive frequency hopping + CRC check anti-interference protocol with a 100ms frequency hopping interval and 16 selectable channels. It automatically avoids electromagnetic interference from high-voltage lines and motor equipment in agricultural and pastoral areas. Employing the LZ77 data compression algorithm, the data size per packet is reduced from 128 bytes to 48 bytes, lowering transmission bandwidth requirements. It also supports the MQTT communication protocol, enabling access to the power grid operation and maintenance cloud platform. Continuous 24-hour transmission interruption time is ≤10s, and the communication success rate is ≥97.8%.

[0191] ③ Signal enhancement unit: Integrated low-noise amplifier to improve the received signal strength in areas with weak signal; equipped with signal detection circuit to monitor signal quality in real time;

[0192] ④ Low-power control unit: adopts intermittent wake-up mode, with a working current of 100mA during data transmission and ≤10μA during sleep mode. It is equipped with a 1000mAh lithium battery backup power supply, which can maintain the operation of the communication module for ≥4 hours when the main power is interrupted, ensuring timely uploading of fault information.

[0193] (4) Power distribution line module, which undertakes the functions of power transmission and segmented reactive power compensation, is adapted to the needs of long-distance and complex environment laying, and consists of line body unit, segmented reactive power compensation unit, status monitoring unit and protection unit:

[0194] ① Line body unit: 95mm² composite copper-clad aluminum core conductor is used. The conductor structure is 19 strands twisted together, the copper layer thickness is 0.2mm, the conductivity is close to that of copper core conductor, and the cost is reduced by 35% compared with pure copper conductor; the insulation layer is made of cross-linked polyethylene (XLPE), with a temperature resistance range of -40℃~90℃ and excellent crack resistance; when buried underground, PE pipe is used for protection, the laying depth is 1.2m, and the joints are made of crimped wire clamps and wrapped with waterproof tape;

[0195] ② Segmented Reactive Power Compensation Unit: Five segmented static var generators (SVG) are deployed at 600m intervals along the 3km line. Each unit has a capacity of 500kvar, a compensation range of -500kvar to +500kvar, and a response time of ≤5ms. The SVG adopts a three-level topology, uses 650V / 300A IGBTs, and has overcurrent, overvoltage, and overtemperature protection functions. It communicates with the intelligent control module via RS485 bus to receive reactive power compensation commands and compensate for line impedance loss in real time.

[0196] ③ Status monitoring unit: It integrates a fiber optic distributed temperature measurement (DTS) system and an impedance real-time calculation module. The DTS sensor is laid along the conductor, with a measurement range of -50℃ to 150℃, a spatial resolution of 1m, and a temperature accuracy of ±0.5℃. It can monitor the temperature distribution of the line in real time. The impedance real-time calculation module calculates the line impedance value by collecting voltage and current signals at the beginning and end of the line, providing data support for intelligent control.

[0197] ④ Protection Unit: The line towers are made of galvanized angle steel, with a height of 8m. The foundation is made of C30 concrete. In areas with frequent sandstorms, the bottom of the towers is equipped with sandstorm protection covers. The conductor suspension points are equipped with anti-vibration hammers to reduce wind vibration damage to the conductors.

[0198] (5) Data processing and feedback module, responsible for data calculation, storage and status feedback, providing decision support for intelligent control, consists of a data processing unit, a storage unit, a calculation optimization unit and a feedback output unit:

[0199] ① Data processing unit: Adopts an industrial-grade ARM Cortex-A9 processor with a main frequency of 1.4GHz, supports floating-point operations, and has a data processing latency of ≤15ms; integrates a 16-bit ADC acquisition module with a sampling rate of 1MHz, capable of acquiring 16 analog signals such as voltage, current, temperature, and dust concentration, with an acquisition accuracy of ±0.1%;

[0200] ② Storage unit: Equipped with a 32GB industrial-grade SD card, with read / write speed ≥80MB / s, storage integrity rate ≥99.9%, and supports circular storage; it also integrates 128MB DDR3 memory for temporary data caching; and is equipped with a power-loss protection circuit, which can support 30s data backup when the main power is interrupted to prevent data loss.

[0201] ③ Operational Optimization Unit: This unit incorporates an adaptive fuzzy PID algorithm and line parameter optimization logic. The adaptive fuzzy PID algorithm includes three stages: fuzzification, fuzzy inference, and defuzzification. The fuzzy domain is {-3,-2,-1,0,1,2,3}, and the membership function uses a triangular function. The control deviation is ≤±1.2%. The line parameter optimization logic calculates the optimal reactive power compensation and voltage regulation based on real-time load and environmental data to ensure minimal line loss.

[0202] ④ Feedback Output Unit: Outputs control commands and operating data via RS485 and Ethernet interfaces, supports Modbus-RTU and TCP / IP protocols; equipped with a 128×64 dot matrix LCD display screen to display key parameters in real time, and also has audible and visual alarm functions.

[0203] (6) The intelligent control module, as the core control center of the device, realizes multi-module coordinated scheduling and consists of a main control unit, a load identification unit, an environmental sensing unit, and an instruction generation unit:

[0204] ① Main control unit: Adopts industrial-grade PLC, with 14 digital inputs, 10 digital outputs, 2 analog inputs, and 1 analog output. It supports PROFINET communication protocol and can simultaneously connect to rectifier, inverter, communication, and data processing modules. The multi-terminal energy regulation coordination rate is ≥99.7%.

[0205] ② Load identification unit: Receives data from the multi-source load monitoring module, uses a wavelet transform-based load identification algorithm to extract load feature vectors, and compares them with a preset load feature library. The load identification accuracy is ≥98%.

[0206] ③Environmental sensing unit: Connects to temperature sensor, dust concentration sensor, and altitude sensor to collect environmental parameters in real time. When the environmental parameters exceed the threshold, an environmental adaptation command is triggered.

[0207] ④ Command generation unit: Based on the "environmental parameters - load characteristics - control strategy" mapping model, it generates three types of commands: first, power conversion command; second, reactive power compensation command; and third, environmental adaptation command. The command generation cycle is ≤100ms to ensure synchronous response of each module.

[0208] (7) Environmental adaptation module, which improves the reliability of the device in extreme environments, consists of a low-temperature heating unit, a sand and dust protection unit, and a high-altitude adaptation unit:

[0209] ① Low-temperature heating unit: A 200W carbon fiber heating element is deployed inside the rectifier and inverter module cabinet and equipped with a temperature controller. The heating will start automatically when the internal temperature of the cabinet is ≤-5℃ and stop when the temperature rises to 5℃. The heating power can be adaptively adjusted according to the temperature deviation.

[0210] ② Dust Protection Unit: The unit cabinet has an IP66 protection rating. The air inlet is equipped with a self-cleaning filter and a filter clogging sensor. When the filter clogging rate is ≥30%, the compressed air self-cleaning function is activated to reduce the filter clogging rate to below 10%.

[0211] ③ High-altitude adaptation unit: For environments above 3000m, the heat dissipation design is optimized, the heat sink area is increased by 15%, and the IGBT drive voltage is adjusted to compensate for the decrease in insulation strength caused by low air pressure at high altitudes, ensuring that the insulation performance of the equipment meets the requirements of GB / T 16935.1-2008.

[0212] 2. Work Steps

[0213] The workflow of this device is divided into seven stages: initialization, data acquisition, data transmission, instruction generation, execution and control, status feedback, and optimization iteration. The specific steps are as follows:

[0214] Step 1: Device Initialization (System Startup Phase)

[0215] Sub-step 1.1: When the main power is turned on, the intelligent control module first performs a self-test, checking the power supply voltage of the rectifier, inverter, communication, and data processing modules, the communication link, and the working status of key components. The self-test takes ≤30 seconds. If a fault is found, an audible and visual alarm is immediately triggered and a fault code is uploaded.

[0216] Sub-step 1.2: After the self-inspection is qualified, the environmental adaptation module starts to collect environmental parameters. The temperature sensor collects the internal temperature of the cabinet and the dust concentration sensor collects the external dust concentration. If the temperature is ≤-5℃, the low temperature heating unit is started; if the dust concentration is ≥1.2g / m³, the filter screen is checked for blockage and the self-cleaning function is started if necessary.

[0217] Sub-step 1.3: The rectifier master module and the inverter slave module initialize their parameters. The rectifier module presets the DC side voltage to 400V, the inverter module presets the output voltage to 220V, and the segmented SVG presets the compensation amount to 0kvar. After initialization, the module enters standby mode and waits for the load to be connected.

[0218] Step 2: Multi-source data collection (real-time monitoring phase)

[0219] Sub-step 2.1: Load data acquisition. The multi-load monitoring module acquires load voltage, current, power factor and power fluctuation data in real time, with an acquisition interval of 100ms. Different acquisition strategies are adopted for different load types: peak starting current acquisition is added for irrigation pump loads, and power periodic change acquisition is added for temperature control equipment.

[0220] Sub-step 2.2: Line data acquisition. The DTS system of the power distribution line module acquires the temperature of each node of the line, the real-time impedance calculation module acquires the voltage and current signals at the beginning and end of the line, and calculates the line impedance value; the segmented SVG acquires its own output current, voltage and compensation data.

[0221] Sub-step 2.3: Environmental data acquisition. The temperature sensor, dust concentration sensor, and altitude sensor of the environmental adaptation module collect environmental parameters, and at the same time, collect the temperature of key components inside the device.

[0222] Sub-step 2.4: Data preprocessing. Each acquisition module preprocesses the raw data: Kalman filtering algorithm is used to remove noise from voltage and current signals, 3σ criterion is used to remove outliers, and the preprocessed data is encapsulated into standard data frames.

[0223] Step 3: Remote Data Transmission (Data Interaction Phase)

[0224] Sub-step 3.1: Data compression. After receiving the data frames sent by each acquisition module, the communication module uses the LZ77 compression algorithm to compress the data: consecutive and repeated data segments are marked and replaced with the original data in the format of "offset + length + data". The data size of a single packet is reduced from 128 bytes to 48 bytes, and the compression time is ≤1ms.

[0225] Sub-step 3.2: Anti-interference transmission. The communication module uses an adaptive frequency hopping protocol to select the communication channel: First, it scans the interference strength of 16 candidate channels and selects the channel with SNR ≥ 10dB as the transmission channel; if the SNR of all channels is < 10dB, the power boosting mechanism is activated; during data transmission, a CRC-16 checksum is embedded, and the receiving end verifies the checksum. If the verification fails, a retransmission is requested.

[0226] Sub-step 3.3: Data distribution. The communication module divides the compressed data packets into two categories: one is real-time control data, which is preferentially transmitted to the data processing and feedback module through the NB-IoT network; the other is historical statistical data, which is transmitted to the power grid operation and maintenance cloud platform during idle periods, with a data transmission accuracy of ≥99.7% at 3km.

[0227] Step 4: Generation of intelligent control instructions (decision-making stage)

[0228] Sub-step 4.1: Data parsing and computation. After receiving the data packet, the data processing and feedback module first decompresses it, and then performs computation through the ARM Cortex-A9 processor.

[0229] a) Load characteristic calculation: Extract load feature vectors based on wavelet transform, compare them with the load feature database, identify the load type, and calculate the load power change rate (ΔP / Δt). If ΔP / Δt ≥ 50 kW / s, it is determined to be a load change.

[0230] b) Line parameter calculation: Based on the voltage and current data at the beginning and end of the line, the line impedance value is calculated using the impedance voltage divider method (R=ΔU / I). Combined with the line temperature measured by DTS, the impedance value is corrected (considering the effect of temperature on conductor resistance: Rt=R20 [1+α(t-20)], where α is the temperature coefficient of aluminum conductor 0.00429 / ℃).

[0231] c) Environmental Impact Calculation: Calculate the environmental impact coefficient based on environmental parameters. For example, in low-temperature environments (t≤-10℃), a 10% increase in conductor resistance necessitates an increase in voltage regulation; when dust concentration ≥1.5g / m³, equipment heat dissipation efficiency decreases by 15%, requiring a 10% reduction in IGBT rated current.

[0232] Sub-step 4.2: Control strategy matching. Based on the calculation results, the intelligent control module calls the "environmental parameters - load characteristics - control strategy" mapping model to generate specific control instructions.

[0233] a) Load change scenario: Trigger the adaptive fuzzy PID algorithm to calculate the DC side voltage regulation of the rectifier module and the reactive power compensation of the segmented SVG, and set the command priority to the highest.

[0234] b) Abnormal environmental scenarios: Generate environmental adaptation instructions, start the rectifier / inverter module heating unit, start the filter self-cleaning function, and set the instruction priority to medium;

[0235] c) Stable operation scenario: PI control algorithm is adopted to finely adjust the voltage of the current adjustment module and the SVG compensation amount to maintain the stability of the terminal voltage, and the command priority is set to the lowest.

[0236] Sub-step 4.3: Instruction verification and encapsulation. The data processing and feedback module verifies the generated control instructions: it checks whether the instruction parameters are within the safe range, and if they exceed the range, it corrects them to safe values; after the verification is successful, the instructions are encapsulated into standard instruction frames.

[0237] Step 5: Execution of Power Conversion and Regulation (Execution Phase)

[0238] Sub-step 5.1: Rectifier module voltage adjustment. After receiving the command, the rectifier host module controls and adjusts the IGBT conduction angle.

[0239] a) If the instruction requires an increase in the DC-side voltage, the STM32F407 microcontroller increases the SVPWM duty cycle and extends the IGBT on-time, increasing the DC-side voltage from 400V to 418V. During the adjustment process, the DC-side voltage is monitored in real time to ensure voltage stability accuracy of ±1.0%.

[0240] b) If the line impedance increases, a fuzzy compensation algorithm is implemented to add an additional voltage regulation amount (ΔU=I×ΔR) based on the impedance change ΔR, in order to offset the voltage drop caused by the increase in impedance.

[0241] Sub-step 5.2: Segmented reactive power compensation. After receiving the command, the segmented SVG adjusts the output reactive power:

[0242] a) The SVG control unit calculates the IGBT switching timing according to the instruction compensation amount and outputs reactive current through a three-level topology; for example, when compensating 70kvar, the output reactive current Iq=70kvar / (√3×220V)=186A;

[0243] b) A current sharing control strategy is adopted, and the compensation amount of the 5 SVGs is distributed according to the line impedance ratio to avoid overload of a single SVG. The compensation response time is ≤5ms.

[0244] Sub-step 5.3: Inverter module voltage regulation. After receiving DC power, the inverter slave module converts it into AC power through a full-bridge topology.

[0245] a) The control and voltage regulation unit adopts SVPWM + third harmonic injection technology to adjust the IGBT switching frequency and conduction angle to stabilize the output AC voltage at 215-220V; when the terminal voltage is low, the modulation amplitude is increased to boost the output voltage.

[0246] b) The harmonic suppression unit monitors the output current harmonics in real time and injects reverse harmonic current through the APF function to control THD to ≤2.9%; in low-temperature environments (t≤-10℃), the heating element of the low-temperature adapter unit is activated to ensure that the IGBT on-state voltage drop is stable at 1.8V and the inverter efficiency is ≥96.1%;

[0247] Sub-step 5.4: Environment adaptation execution. After receiving the instruction, the environment adaptation module starts the corresponding function:

[0248] a) Low-temperature heating: The carbon fiber heating element is powered on and heated. The temperature controller monitors the cabinet temperature in real time and stops heating when it reaches 5°C to avoid excessive energy consumption.

[0249] b) Filter self-cleaning: Open the compressed air valve and flush the filter with 0.6MPa compressed air. Monitor the filter pressure difference during the cleaning process. Stop cleaning when the pressure difference drops below 50Pa.

[0250] Step 6: Operational Status Feedback (Monitoring and Early Warning Phase)

[0251] Sub-step 6.1: Status data acquisition. Each module acquires its own operating status data in real time: the rectifier module acquires IGBT junction temperature, DC side voltage, and output current; the inverter module acquires output voltage THD and AC side current; the communication module acquires signal strength and transmission success rate; the SVG acquires compensation current and power factor; the acquisition interval is 1 second.

[0252] Sub-step 6.2: Anomaly detection and early warning; the data processing and feedback module performs threshold judgment on the status data:

[0253] a) If the line temperature is ≥70℃, the fiber optic distributed temperature measurement module will be triggered to alarm, and an optimization instruction will be generated at the same time: adjust the SVG compensation amount, reduce the line current, and reduce the line loss.

[0254] b) If the terminal voltage deviation is > ±2%, increase the DC side voltage regulation of the rectifier module to ensure that the voltage returns to the 215-220V range;

[0255] c) If the equipment fails, immediately generate a fault command: cut off the power supply to the corresponding module, upload the fault code to the operation and maintenance cloud platform, and start the backup module at the same time;

[0256] Sub-step 6.3: Operation report generation. The data processing and feedback module automatically generates an operation report daily (00:00), which includes the following content:

[0257] a) Power quality indicators: end-point voltage qualification rate (≥99.5%), line loss rate (daily average ≤4.5%), THD (daily average ≤2.9%).

[0258] b) Equipment operation indicators: fault-free running time, startup success rate, and maintenance frequency of each module;

[0259] c) Environmental adaptability indicators: low-temperature heating time, number of times the filter screen is cleaned, and number of times the dust concentration exceeds the standard;

[0260] d) The report is in PDF format and is pushed to the power grid operation and maintenance platform via the communication module, while also being stored on an SD card for backup.

[0261] Step 7: Optimization and Iteration (Continuous Improvement Phase)

[0262] Sub-step 7.1: Data statistical analysis. After receiving operation reports from multiple devices, the power grid operation and maintenance platform performs statistical analysis: average line loss rate and equipment failure rate by region; adaptation success rate by load type; and environmental adaptation effect by environmental conditions.

[0263] Sub-step 7.2: Strategy optimization. Based on the statistical analysis results, optimize the mapping model of "environmental parameters - load characteristics - control strategy". For example, for the plateau environment of Hainan Prefecture, Qinghai, adjust the voltage regulation coefficient to compensate for the voltage drop caused by the low air pressure on the plateau; for the severe winter in Altay, Xinjiang, extend the preheating time of the heating element.

[0264] Sub-step 7.3: Firmware update. The optimized algorithm firmware is pushed to the device via the NB-IoT network. After receiving the firmware, the smart control module verifies it. If the verification is successful, the update is automatically performed. The update process does not affect the normal operation of the device.

[0265] Based on the same inventive concept, this application also provides an intelligent low-voltage flexible DC-environment adaptive multi-energy control device for implementing the above-mentioned intelligent low-voltage flexible DC-environment adaptive multi-energy control device method. For example... Figure 3 As shown, it includes: a distribution network data acquisition module, an energy demand adaptation module, an energy state transition module, and an operation strategy optimization module. The solution provided by this device is similar to the solution described in the above method.

[0266] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0267] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0268] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0269] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0270] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0271] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0272] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0273] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A smart low-voltage flexible direct current - environment adaptive multi-purpose energy utilization control method, characterized in that, The method comprises: obtaining distribution network load data, distribution network line data and distribution network environment data of a power system; performing load characteristic operation on the distribution network load data to obtain distribution network load characteristic information; performing line parameter operation on the distribution network line data to obtain distribution network line characteristic information; performing environmental impact operation on the distribution network environment data to obtain distribution network environment characteristic information; performing counterfactual causal correction on the distribution network load characteristic information according to the distribution network environment characteristic information to obtain load counterfactual characteristic information; performing physical constraint graph homotopy continuous processing on impedance and topology data in the distribution network line characteristic information to obtain line load coupling manifold parameters; performing weighted least square evaluation on each power parameter of the power system according to the load counterfactual characteristic information and the line load coupling manifold parameters to obtain a target evaluation vector; performing dual original collaborative solution of a differentiable mixed integer programming on the power system according to the target evaluation vector to obtain energy demand adaptation analysis data; performing dynamic energy demand adaptation mapping on the power system according to the energy demand adaptation analysis data to obtain energy demand adaptation mapping instructions; controlling the power system to perform electric energy state conversion according to the energy demand adaptation mapping instructions to obtain distribution network operation state data; performing time sequence comparison learning representation processing on the distribution network operation state data to obtain an operation situation vector; performing topology sensitivity sparsification analysis on adjustable nodes in the power system according to the operation situation vector to obtain an adjustable node set; performing weighted least square correction on voltage and reactive power measurement data in the adjustable node set to obtain an initial parameter adjustment amount; performing Wasserstein ball distribution modeling on a disturbance set in a prediction model of the differentiable robust model predictive control solution according to the operation situation vector and the initial parameter adjustment amount to obtain a robust uncertainty set parameter; performing soft constraint weighted least square setting on a prediction cost function of the differentiable robust model predictive control solution to obtain a soft constraint weight and an initial Lagrange multiplier; constructing model predictive control information of the prediction model in a quadratic programming form according to the robust uncertainty set parameter and the soft constraint weight to obtain a quadratic programming controller; performing implicit layering differentiable parameterization processing on the quadratic programming controller to obtain a differentiable quadratic programming controller; performing dual disturbance alignment solution in a time domain rolling window according to the differentiable quadratic programming controller to obtain distribution network energy dynamic control data.

2. The method of claim 1, wherein, The method comprises: performing multi-element energy scene analysis on the power system according to the energy demand adaptation analysis data to obtain each scene energy analysis data; performing dynamic energy demand strategy analysis on the power system according to each scene energy analysis data to obtain each energy parameter strategy analysis data; performing strategy fusion mapping on each energy parameter strategy analysis data to obtain the energy demand adaptation mapping instructions.

3. The method of claim 2, wherein, The use energy analysis data of each scene is subjected to counterfactual disturbance analysis to obtain a scene counterfactual sample set. According to the scene counterfactual sample set, extreme value optimization within a distribution feasible region is performed on each candidate use energy parameter of the power system to obtain a candidate use energy parameter set. The candidate use energy parameter set is subjected to weighted least square consistency checking to obtain a compliance candidate parameter set. According to the compliance candidate parameter set and the scene counterfactual sample set, a differentiable mixed integer strategy distillation mapping is performed on a strategy space of the power system to obtain each use energy parameter strategy analysis data. The use energy parameter strategy analysis data is subjected to strategy fusion mapping to obtain the use energy demand adaptation mapping instruction, including:

4. The method of claim 2, wherein, The use energy parameter strategy analysis data is subjected to dual embedding representation processing to obtain a strategy dual representation set. According to the strategy dual representation set, zero-knowledge game analysis is performed on conflicting strategies in each use energy parameter strategy analysis data to obtain a mutually compatible strategy subset. The mutually compatible strategy subset is subjected to weighted voting fusion to obtain a preliminary fusion strategy parameter. The preliminary fusion strategy parameter is subjected to neural satisfiability driven strategy synthesis to obtain the use energy demand adaptation mapping instruction. The device includes:

5. A smart low-voltage flexible direct current - environment adaptive multi-purpose energy control device, characterized in that, A distribution network data acquisition module for acquiring distribution network load data, distribution network line data, and distribution network environment data of a power system; A use energy demand adaptation module for performing load characteristic operation on the distribution network load data to obtain distribution network load feature information; Performing line parameter operation on the distribution network line data to obtain distribution network line feature information; Performing environmental impact operation on the distribution network environment data to obtain distribution network environment feature information; According to the distribution network environment feature information, counterfactual causal correction is performed on the distribution network load feature information to obtain load counterfactual feature information; The impedance and topology data in the distribution network line feature information are subjected to physical constraint graph homotopy continuous processing to obtain line load coupling manifold parameters; According to the load counterfactual feature information and the line load coupling manifold parameters, weighted least square evaluation is performed on each power parameter of the power system to obtain a target evaluation vector; According to the target evaluation vector, dual original collaborative solution of differentiable mixed integer programming is performed on the power system to obtain use energy demand adaptation analysis data; The use energy demand adaptation module is further configured to perform dynamic use energy demand adaptation mapping on the power system according to the use energy demand adaptation analysis data to obtain use energy demand adaptation mapping instructions; An electric energy state conversion module for controlling the power system to perform electric energy state conversion according to the use energy demand adaptation mapping instructions to acquire distribution network operation state data; An operation strategy optimization module for performing time series comparative learning representation processing on the distribution network operation state data to obtain an operation situation vector; ​ According to the operation situation vector, a topological sensitivity sparsification analysis is performed on an adjustable node in the power system to obtain an adjustable node set; Weighted least square correction is performed on voltage and reactive power measurement data in the adjustable node set to obtain an initial parameter adjustment amount; According to the operation situation vector and the initial parameter adjustment amount, a Wassertein ball distribution modeling is performed on a disturbance set in a prediction model for solving the differentiable robust model predictive control to obtain a robust uncertainty set parameter; Soft constraint weighted least square setting is performed on a prediction cost function for solving the differentiable robust model predictive control to obtain a soft constraint weight and an initial Lagrange multiplier; According to the robust uncertainty set parameter and the soft constraint weight, a model predictive control information of the prediction model is constructed in a quadratic programming form to obtain a quadratic programming controller; A differentiable parameterization processing is performed on the quadratic programming controller in an implicit layering manner to obtain a differentiable quadratic programming controller; According to the differentiable quadratic programming controller, a dual disturbance alignment solving is performed in a time domain rolling window to obtain distribution network energy utilization dynamic control data. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-4 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 4.

Images