A mine scene-oriented micro-grid multi-scene cooperative optimization scheduling method

Abstract

The present application relates to the technical field of industrial energy system optimal scheduling and intelligent control, and proposes a micro-grid multi-scenario collaborative optimization scheduling method for mine scene, which comprises the following steps: data preprocessing and system basic modeling for mine characteristics, construction of fine benchmark economic optimal scheduling model, emergency standby scheduling model based on distribution robust opportunity constraint, and collaborative optimization model of virtual power plant participating in day-ahead spot market, adoption of multi-mode adaptive switching mechanism based on rolling time domain and state discrimination to optimize model parameters, and finally integration of each module by using micro-service architecture and deployment in mine scheduling center to realize uninterrupted automatic optimization scheduling.The present application realizes the global collaboration of "economic in normal state, safety in emergency, and market in opportunity" in mine micro-grid for the first time by constructing three independent and complete optimization models of economy, emergency and market, and designing intelligent switching mechanism in the upper layer.

Description

Technical Field

[0001] This invention relates to the field of industrial energy system optimization scheduling and intelligent control technology, and more specifically, to a multi-scenario collaborative optimization scheduling method for microgrids in mining scenarios. Background Technology

[0002] As an energy-intensive industry, mining is undergoing a transformation in its energy systems, shifting from reliance on traditional fossil fuels and high-carbon power grids to microgrid models that integrate clean energy. Mining microgrids aim to reduce energy costs, decrease carbon emissions, and enhance power supply autonomy by integrating local distributed renewable energy sources such as solar and wind power, configuring energy storage systems, and optimizing the scheduling of traditional power sources like diesel generators. However, the unique characteristics of the mining production environment—including complex load structures, extremely high safety and reliability requirements, and harsh natural conditions—present significantly greater technical challenges to the operation and management of these microgrids than those faced by ordinary commercial or residential microgrids.

[0003] Existing technologies have yielded extensive research on the optimal scheduling of microgrids, primarily falling into three categories. However, none of these approaches have systematically addressed the multi-dimensional coordination challenges in mining scenarios. Single-objective economic dispatch research: This type of research focuses on minimizing the operating cost of microgrids and typically constructs deterministic or stochastic optimization models. Common methods include establishing mixed-integer linear programming (MILP) or mixed-integer second-order cone programming (MISOCP) models to coordinate the dispatch of photovoltaic, wind turbines, energy storage, diesel generators, and loads. For example, a typical model aims to minimize the sum of electricity purchase cost, fuel cost, and equipment maintenance cost, with constraints including power balance, energy storage charging and discharging dynamics, generator operating ranges, and ramp-up conditions. However, this type of research generally treats the total load within the microgrid as an aggregated, predictable variable, or simply categorizes it. For mines, this approach is severely distorted because it ignores the rigidity, high power, and strong coupling between productive loads and production plans, and fails to distinguish between critical loads that must be absolutely guaranteed and flexible loads with adjustment potential. Therefore, dispatch schemes based on this coarse-grained model, while theoretically reducing some costs, cannot fully exploit the huge flexibility potential on the mine load side, resulting in limited economic improvement and potentially rendering the scheme infeasible in practice due to model distortion.

[0004] Research on Emergency Power Supply for Reliability: To address the risk of power outages, existing technologies primarily rely on the "N-1" or "N-2" safety criteria for redundant equipment configuration and, upon detecting a fault, switchover based on pre-defined logical rules (e.g., "if the main power supply fails, all backup diesel generators will be activated"). More advanced methods employ robust optimization or chance-constrained programming to address uncertainty. Robust optimization assumes that uncertainty varies within a defined set and optimizes performance under worst-case conditions, but its results are often overly conservative, leading to excessively large backup capacity and poor economic efficiency. Chance-constrained programming allows constraints to be violated with a certain probability, but requires accurate knowledge of the probability distribution of uncertainty, which is difficult to obtain in mining fault scenarios. Existing solutions fail to use technologies such as digital twins to scenario-based and data-driven modeling of mine-specific fault modes, and also lack an optimization modeling method that can balance reliability and economy with limited fault sample data.

[0005] Research on Virtual Power Plant Participation in the Electricity Market: With the development of the electricity market, numerous studies have focused on the bidding strategies of microgrids as virtual power plants participating in the energy and ancillary service markets. One mainstream approach is two-stage stochastic programming: the first stage determines the day-ahead market bid volume, and the second stage optimizes internal resources to minimize expected operating costs or maximize revenue for various renewable energy and price scenarios. Another approach is biblical optimization, which does not assume a specific probability distribution but considers the worst-case scenario within a fuzzy set. However, most of these studies assume that the load is a simple linear or piecewise linear function with known price elasticity. For mines, the response behavior of their adjustable load (mainly electric mining truck charging) is highly nonlinear and influenced by multiple complex factors such as production scheduling, vehicle operation cycles, driver behavior, and battery status, making it difficult to accurately describe using traditional mathematical models. Directly applying existing market models can lead to bidding strategies deviating from actual adjustment capabilities, either missing arbitrage opportunities or generating unbalanced deviations in real-time operation, resulting in financial risks.

[0006] Existing technology proposes a scenario-based two-stage stochastic programming method for microgrid market bidding and dispatch. This method first uses historical data to generate a set of discrete scenarios and their probabilities regarding the next day's renewable energy output and market prices. In the first stage (day-ahead), the optimization model determines the microgrid's power purchase and sale bidding curve and generator start-up and shutdown plan in the day-ahead market. In the second stage (intra-day), for each scenario, the optimization model dispatches internal resources such as energy storage and adjustable loads to minimize the total cost (including day-ahead market settlement costs and real-time balancing costs) under that scenario. The ultimate goal is to minimize the expected total cost across all scenarios.

[0007] The shortcomings of this approximation scheme are: The model is fragmented: it is designed only for market participation models and does not include a dedicated, highly reliable emergency dispatch model. When the system fails, there is a lack of an optimization mechanism that is linked to it and has the core objective of ensuring system security.

[0008] Load modeling is simplistic: adjustable loads are typically modeled as a linear response function to electricity prices, or given fixed adjustable upper and lower limits. This completely fails to capture the complex dynamic characteristics of flexible loads in mines, resulting in a significant gap between dispatch instructions and actual responses.

[0009] Limitations in handling uncertainty: The effectiveness of scenario stochastic planning heavily depends on the quality and quantity of the generated scenarios. In environments like mines where data is relatively scarce and the sources of uncertainty (equipment failures) have a "long-tail distribution" characteristic, it is difficult to accurately capture all risks through a limited number of scenarios, especially extremely low-probability, high-loss events.

[0010] Lack of adaptive capability: This is an offline, day-ahead decision-making model that cannot dynamically adjust its operating strategy and mode based on real-time operating status (such as sudden equipment failure or drastic market price fluctuations) within the day.

[0011] Therefore, although this approximation scheme theoretically provides a framework for handling market uncertainty, it cannot meet the complex requirements of efficient, safe, and flexible operation of mining microgrids in terms of model completeness, scenario adaptability, decision-making precision, and system resilience. Summary of the Invention

[0012] In view of this, the present invention proposes a multi-scenario collaborative optimization scheduling method for microgrids in mining scenarios. Physically, the load is divided into four categories: productive, critical guarantee, adjustable, and basic. A mathematical model that conforms to the operating characteristics of each category is established, especially the flexibility boundary model of adjustable loads, thereby elevating the scheduling optimization from "total control" to the level of "classified precise regulation".

[0013] To achieve the above objectives, the present invention proposes the following technical solution: A multi-scenario collaborative optimization scheduling method for microgrids in mining scenarios includes: Step 1: Data preprocessing and basic system modeling tailored to the characteristics of the mine: S1.1 Establish a unified data acquisition platform to synchronously collect power data, load data, environmental and market data at a frequency of no less than 1 minute; adopt the "3σ criterion" in combination with the upper and lower limits of equipment operation to identify and remove outliers; use time series prediction model imputation or decision tree-based method to fill missing data; and use wavelet transform for smoothing filtering of photovoltaic data. S1.2 Load Refinement Classification Modeling Based on Physical Mechanisms and Data-Driven Approach: Decomposing the total load into four independent variables: Productivity load: Establish a mathematical model based on equipment start-up and shutdown status, rated power, and load rate, and set constraints on minimum continuous operation time, minimum downtime, and workload requirements; Critical load protection: Establish a mathematical model of rated power and allowable fluctuations as a hard constraint to ensure power supply continuity; Adjustable load: Establish a flexible mathematical model that includes baseline planning and up / down power adjustment, and set dynamic boundary constraints and energy conservation constraints, wherein the adjustable power boundary is dynamically output by the data-driven model; Basic load: Forecasted using a time series model; S1.3 LSTM neural network combined with production calendar is used for load forecasting, physical-statistical hybrid model combined with numerical weather prediction is used for photovoltaic power output forecasting, power conversion model based on wind speed forecasting is used for wind power output forecasting, and GARCH model or deep probabilistic forecasting model is used for electricity price forecasting. Step Two: Constructing Three Core Optimization Scheduling Models: S2.1 With minimizing the total operating cost within the scheduling cycle as the objective function, a mixed integer linear programming model is established, which includes refined power balance constraints, detailed energy storage model, diesel generator start-stop logic constraints, power grid interaction constraints, and adjustable load constraints, to construct a refined benchmark economic optimal scheduling model; S2.2 Based on the mine digital twin system, an emergency scenario library is constructed. The model is modeled using the joint chance constraint of the sub-Brussels bar based on Wasserstein distance. The probabilistic constraints are equivalently transformed into deterministic linear constraints. With the goal of minimizing emergency fuel costs or maximizing safe time periods, an emergency standby scheduling model based on the chance constraint of the sub-Brussels bar is constructed. S2.3 Establish a two-stage decision-making framework. The first stage determines the day-ahead market bidding volume and generator start-up and shutdown status. The second stage determines the real-time scheduling of internal resources in response to the uncertainties of renewable energy output and real-time electricity prices. A machine learning model is used to model the dynamic feasible region of adjustable loads. A hybrid stochastic / robust optimization is used to handle the dual uncertainties. The two-stage problem is reconstructed into a single-stage mixed integer linear programming model. A collaborative optimization model for virtual power plants participating in the day-ahead spot market is constructed. Step 3: Multi-mode adaptive switching mechanism based on rolling time domain and state discrimination: S3.1 Trigger the scheduling cycle at fixed time intervals, obtain the latest state measurement values ​​and prediction data, call the corresponding optimization model to solve, only use the control instructions for the current time period to be issued and executed, and wait for the next cycle to repeat the above process; S3.2 The emergency mode priority criterion and the market mode opportunity criterion are executed sequentially to determine the current operating mode based on the equipment fault signal, the fault probability prediction value, the expected arbitrage index, and the system flexibility status. S3.3 Smooth transition during execution mode switching is ensured by a combination of strategies including state inheritance, preservation of common constraints, softening of constraints during transition periods, and filtering of output instructions to prevent sudden changes in control instructions during switching. S3.4 The prediction model is updated using an online learning algorithm, a closed-loop data flow is established to iteratively update the data-driven model, and the optimization parameters are dynamically adjusted based on the running performance. Step 4: Integrate the various modules using a microservice architecture and deploy them in the mine scheduling center to achieve uninterrupted automatic optimization scheduling.

[0014] Furthermore, the dynamic boundary constraints and energy conservation constraints in the adjustable load of S1.2 are as follows: Dynamic boundary constraints: ; Energy conservation constraint: ; in, and Data-driven model The system dynamically outputs data based on market characteristics, production characteristics, time characteristics, vehicle status characteristics, and historical response characteristics.

[0015] Furthermore, the joint chance constraint modeling of the bibru bars based on Wasserstein distance in S2.2 specifically includes: Define the power supply margin function at time t:

[0016] in For emergency dispatch decision variables; Construct joint opportunity constraints: ; Among them, fuzzy set Defined as: , The distance is 1-Wasserstein. Let be the radius.

[0017] Furthermore, the bibliometric joint chance constraint is transformed into a mixed integer linear programming constraint set by introducing auxiliary variables, thus completely transforming the complex probabilistic constraints into deterministic linear and integer constraints.

[0018] Furthermore, the data-driven dynamic feasible domain modeling of adjustable load described in S2.3 specifically includes: Training machine learning models Input features include: Market characteristics: Day-ahead electricity prices, real-time electricity price forecasts, and their volatility; Production characteristics: shift schedule, planned number of vehicles dispatched, current shift's charged battery capacity, average transport distance; Time characteristics: moment, weekday / holiday, season; Vehicle status characteristics: average battery SOC, number of vehicles waiting to be charged; Historical response characteristics: actual adjustment amounts under similar scenarios in the past 24 hours; The model outputs the prediction range of adjustable load power for that period. As for this period The upper and lower bounds of the hard constraints.

[0019] Furthermore, the hybrid stochastic / robust optimization handling of dual uncertainties described in S2.3 specifically includes: For renewable energy output, S typical scenarios generated by clustering historical data are used to characterize it; For real-time electricity prices, robust modeling is performed using a budget uncertainty set:

[0020] in, It is a predicted value. It is the magnitude of the prediction error. It is a conservatism parameter, which controls the maximum number of time periods during which the electricity price can simultaneously take the worst value; By utilizing the duality theory of linear programming, the inner-level max problem is transformed into a min problem, and the two-stage problem is reconstructed into a single-stage mixed-integer linear programming model.

[0021] Furthermore, the emergency mode priority criterion described in S3.2 includes: Hard alarm: Directly read the fault hard signal from the equipment monitoring system. If a fault is detected, immediately set the mode flag to "emergency". Soft early warning: Query the equipment health prediction system to obtain information on the future health of critical equipment. Failure probability prediction value within the time period ,like Then set the mode to "emergency".

[0022] Furthermore, the market pattern opportunity criteria described in S3.2 include: Calculate the expected arbitrage index within the future optimization window: :

[0023] in, To optimize the expected arbitrage index within the future window, Indicates the current moment. This indicates an optimization of the window length. express Time weight, express The daytime electricity price express The benchmark electricity price at any given time. This is the risk aversion coefficient, used to balance arbitrage profits with the risk of electricity price fluctuations. The standard deviation of the day-ahead electricity price represents the risk of electricity price volatility. Perform system flexibility status verification to determine whether the energy storage status is within a safe range and whether the adjustable load potential is greater than the threshold. like If the flexibility check passes, the mode is set to "Market"; otherwise, the mode is set to "Economy".

[0024] Furthermore, the smooth transition guarantee for mode switching described in S3.3 includes: State inheritance: In any mode, the initial state of the optimized model is strictly equal to the actual measured state at the end of the previous time period; Common constraints are maintained: the three models share the basic physical and security constraints of all devices; Transition period soft constraint: When it is determined that the mode is about to switch, add an additional soft constraint on the rate of change of power for the productive load; Output command filtering: A first-order low-pass filter is used to smooth the command value, and a smaller filter coefficient is used at the mode switching time.

[0025] Furthermore, the online self-learning and parameter adaptation of the model described in S3.4 includes: Online updates to the prediction model: using recursive parameter estimation of the ARIMA model or online fine-tuning of the neural network; Data-driven model Iteration: Establish a closed-loop data flow, and periodically add actual execution data as new samples to the training set for incremental updates or full retraining; Optimize parameters dynamically: Based on the running performance, dynamically adjust the Wasserstein radius, robust budget, risk tolerance, and mode switching threshold through Bayesian optimization or trial and error.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs three independent and complete optimization models—economic, emergency, and market-based—and designs an intelligent switching mechanism at the upper level. For the first time, this invention achieves global coordination in a mining microgrid, prioritizing "economic efficiency in normal times, safety in emergencies, and market opportunities." The system is no longer a simple aggregation of multiple isolated strategies, but an intelligent agent capable of automatically selecting the optimal posture based on value judgments. This allows the mining microgrid to consistently approach the globally optimal operating point under different scenarios, with comprehensive benefits (economic efficiency + safety + market returns) far exceeding any single-mode or fixed-weighted scheme.

[0027] This invention pioneers a four-type load independent modeling method, especially establishing a refined model of "energy conservation + dynamic power boundary" for adjustable loads. This allows the optimization engine to clearly distinguish between "immovable," "must-protect," and "adjustable" loads. Scheduling instructions have shifted from adjusting the "total load" to "precise drip irrigation" for each type of load. This greatly unleashes the flexibility potential of the load side, increasing the economic scheduling cost reduction from less than 10% to 15%-25%, and the formulated strategies are fully consistent with actual production and highly executable.

[0028] This invention introduces a bibliometric chance-constrained programming approach based on Wasserstein distance, transforming the requirement of "ensuring power supply to critical loads" from a qualitative, fuzzy safety criterion into a precise mathematical constraint of "ensuring power supply under probability." Through parameter summation, operators can precisely and flexibly balance "safety level" and "economic cost," much like adjusting a knob. This achieves the lowest emergency configuration cost at the same safety level, or the highest safety level at the same cost, completely resolving the pain points of traditional methods that either result in excessive waste or risky backups.

[0029] This invention innovatively integrates a data-driven machine learning model with an advanced hybrid stochastic / robust optimization framework. The model enables the optimizer to "know" the true boundary of load adjustment capacity under specific production conditions and market signals. The hybrid optimization framework ensures that, in the face of uncertain market prices and renewable energy, decisions both actively seize opportunities and prudently defend against risks. This transforms the final bidding strategy from mere theoretical speculation into a highly executable, high-yield, and low-risk intelligent decision, shifting mines from passive price takers to proactive market participants with management capabilities.

[0030] This invention, based on online optimization in the rolling time domain and embedded pattern discrimination logic, forms a complete closed loop of "perception-decision-execution-learning". The system can "perceive" changes in the internal and external environment (faults, electricity prices) in real time, automatically "decide" and switch to the optimal operating mode, smoothly "execute" scheduling instructions, and "learn" from operational data to improve its own model. This adaptive capability endows the mine microgrid with strong system resilience, enabling it to cope with various emergencies and complex scenarios, achieving a leap from "automation" to "intelligence", significantly reducing manual intervention, and improving operational efficiency and safety. Attached Figure Description

[0031] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of the microgrid multi-scenario collaborative optimization scheduling method for mining scenarios according to the present invention. Detailed Implementation

[0032] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] like Figure 1 As shown in the figure, this embodiment proposes a complete, closed-loop multi-scenario collaborative optimization scheduling method for mine microgrids. This method consists of four main technical steps, forming a complete chain from data to model, then to decision-making and execution. The following will elaborate on each step in detail.

[0034] Step 1: Refined Data Preprocessing and Basic System Modeling Based on Mine Characteristics This step forms the basis for all subsequent advanced optimization features and is designed to create an accurate and reliable digital image of the system.

[0035] S1.1 Multi-source heterogeneous data acquisition and deep fusion cleaning Mine data comes from diverse sources, including SCADA systems (power generation and distribution), production management systems (MES / APS), equipment health management systems (PHM), environmental monitoring stations, and smart meters. This invention first establishes a unified data acquisition platform, synchronously collecting data at a frequency of no less than one minute. Power data: Real-time output and forecast curves of photovoltaic / wind power; SOC, charging and discharging power, and health status of energy storage; output, operating hours, and fault codes of diesel generators.

[0036] Load data: Intelligent monitoring units are installed at key power distribution nodes to preliminarily decompose the total load curve through load characteristic analysis and pattern recognition. Simultaneously, they interface with the production system to obtain start-up and shutdown plans and real-time power of large equipment such as electric shovels, crushers, and belt conveyors.

[0037] Environmental and market data: irradiance, temperature, wind speed; recently released electricity price curves and real-time price forecasts.

[0038] Data cleaning: The "3σ criterion" is used in conjunction with the upper and lower limits of equipment operation to identify and remove obvious outliers. For missing data, different strategies are adopted for different types: continuous power data is imputed using time series prediction models (such as ARIMA); discrete state data is filled using the mode of the preceding and following time points or a decision tree-based method. For rapid fluctuations in photovoltaic data caused by cloud cover, wavelet transform is used for smoothing filtering to preserve trends and remove noise.

[0039] S1.2 Load Refinement Classification Modeling Based on Physical Mechanisms and Data-Driven Approach This is one of the core foundations of this invention. We will consider the total load. Decompose it into four independent variables with clear physical meaning and mathematical description: Productivity load : Modeling objects include: electric shovels, hydraulic crushers, semi-mobile crushing plants, and large ball mills. The real-time power of the productive load is determined by the equipment's start / stop status, rated power, and load rate, as shown in the following formula: Mathematical model:

[0040] in, For the number of productive load equipment, This indicates the start / stop status of device m during time period t (1 for running, 0 for stopping), and its value is determined by the production plan; Let m be the rated power of the equipment; The load rate (0~1) of equipment m during time period t is related to the hardness of the working material, digging depth, and equipment wear, and can be obtained in real time through historical data statistics or sensors to meet the requirements. ,in and Let m be the minimum and maximum load rates of the device.

[0041] Key constraint: Minimum continuous running time Minimum downtime And the daily / shift workload requirements.

[0042] Critical protection load : Modeling objects: main ventilation fan, underground main drainage pump, emergency rescue lighting, safety monitoring and communication system, gas drainage pump, etc.

[0043] Mathematical model: .

[0044] Rated or base power, typically operating at constant or in a fixed mode; For allowable fluctuations, such as when a ventilator adjusts according to wind speed, to meet... .

[0045] Core requirement: In any optimization model, the following must be satisfied. The continuity of power supply is usually considered a hard constraint.

[0046] Adjustable load : Modeling objects: charging pile groups for electric mining trucks (such as 100-ton class), intermittently operating air compressors, air conditioners in non-core areas, etc.

[0047] Flexibility mathematical model: .

[0048] Baseline charging / operation plan, initially provided by production scheduling; : Upward and downward adjustment power relative to the baseline.

[0049] Dynamic boundary constraints: , .

[0050] Energy conservation constraint: This ensures that the total energy demand within a cycle (such as a shift) is met.

[0051] in, and It is not a fixed value, but rather a data-driven model trained in subsequent steps. Dynamic output is key to achieving precise market response.

[0052] Basic load : Modeling objects: administrative offices, living area lighting and air conditioning, maintenance workshops, etc.

[0053] Processing method: Its power curve is relatively stable and highly predictable. It is predicted using a time series model (such as exponential smoothing with seasonal decomposition). In optimization, it is regarded as a known parameter or a deterministic quantity with small fluctuations.

[0054] S1.3 Integration of High-Precision Short-Term Prediction Models To support rolling optimization, forecast information for the next few hours to 24 hours is required. This invention integrates multiple forecasting modules: Load forecasting: , , The prediction is performed using an LSTM neural network that incorporates the production calendar.

[0055] New energy forecasting: Photovoltaic power output adopts a physical-statistical hybrid model that combines numerical weather prediction; wind power output adopts a power conversion model based on wind speed prediction.

[0056] Electricity price forecast: Day-ahead electricity prices are considered known; real-time electricity price forecasts use GARCH models or deep probabilistic forecasting models, and output point predictions and confidence intervals for subsequent robust optimization.

[0057] Step 2: Construct three core optimization scheduling models Based on the refined model, optimized scheduling models for three different scenarios are constructed respectively.

[0058] S2.1 Mode 1: Refined Benchmark Economic Optimal Scheduling Model (Daily Operation) The goal of this model is to minimize the total operating cost within a scheduling period T (e.g., 24 hours) while satisfying all physical constraints and load requirements of the devices.

[0059] Objective function: , in, It is a 0-1 variable, representing the start-up and shutdown actions of generator g at time t.

[0060] Core constraint system: Refined power balancing: As mentioned above, it is the sum of four types of loads.

[0061] Detailed energy storage model:

[0062] in Self-discharge loss is modeled as a function of SOC.

[0063] Diesel generator combination: In addition to output upper and lower limits and ramping constraints, the start-stop logic is strictly modeled.

[0064] Grid Interaction: .

[0065] This model is a standard mixed-integer linear programming problem that can be solved efficiently using commercial solvers.

[0066] S2.2 Mode 2: Emergency Backup Scheduling Model Based on Divided Bar Chance Constraints When the system enters an emergency state, the objective changes to ensuring power supply to critical loads with the highest probability under extremely uncertain conditions, and extending the protection time as much as possible or minimizing emergency costs.

[0067] Emergency scenario database construction: Based on the mine digital twin system, simulations are performed on typical composite faults such as "single diesel engine failure + cloudy weather", "partial failure of energy storage system", and "external power grid loss + wind turbine power restriction", generating N multi-dimensional time series samples; Constituting the empirical distribution .

[0068] Modeling of joint chance constraints for Brussels-based bars: Define the power supply margin function at time t: .

[0069] in These are variables for emergency dispatch decision-making (energy storage discharge power, standby diesel engine output, etc.).

[0070] The requirement is to calculate the probability throughout the entire emergency period under the most unfavorable probability distribution. The joint probability that the power supply margin is non-negative at all times is not less than :

[0071] Here is a fuzzy set Defined as: , The distance is 1-Wasserstein. Let be the radius.

[0072] Equivalent deterministic reconstruction of the model (key transformation steps): According to DRCC theory, the above constraint can be transformed into a constraint on conditional value of risk. By introducing auxiliary variables, it can be further equivalently transformed into the following mixed-integer linear programming constraint set (for clarity, a single time period is used as an example; the principle is similar for multiple time periods but more auxiliary variables are needed to handle joint probabilities):

[0073] in, It is linearization Auxiliary variables. With decision variables This is related to a certain norm, and can be further simplified under a specific linear mine model. This transformation completely converts complex probabilistic constraints into deterministic linear and integer constraints, making the problem solvable.

[0074] Special strategies and objectives in emergency mode: Objective function: can be set as (Minimize emergency fuel costs) or (Maximize the safe period).

[0075] Policy constraints: mandatory Release all energy storage discharge capacity (the lower limit of SOC can be temporarily lowered), and prioritize the use of high-efficiency standby diesel engines.

[0076] S2.3 Mode 3: Collaborative Optimization Model for Virtual Power Plants Participating in the Day-ahead Spot Market In this model, the mining microgrid participates in the market as a whole, with the goal of maximizing expected net revenue.

[0077] Two-stage decision-making framework: Phase 1 (Day-to-Day): The decision variable is the day-to-day market bid volume. (Purchase positive, sell negative), and start / stop of generator sets. This decision was made just before the market was shut down and is irreversible.

[0078] Phase Two (Real-Time): After the known day-ahead bidding results and unit status, the focus is on renewable energy output. and real-time electricity price Uncertainty in decision-making, real-time scheduling of internal resources (energy storage, adjustable load) , Wait, in order to balance the deviation from the current plan. .

[0079] Objective function – maximizing expected net profit:

[0080] in, It is the cost of real-time deviation settlement. It refers to internal operating costs.

[0081] Data-driven dynamic feasible region modeling of adjustable loads: Training a machine learning model (e.g., XGBoost or LightGBM). Feature engineering is crucial, including: Market characteristics: Day-ahead electricity price Real-time electricity price forecast and its volatility.

[0082] Production characteristics: shift schedule, planned number of vehicles dispatched, current shift's charged charge, average transport distance.

[0083] Time characteristics: moment, weekday / holiday, season.

[0084] Vehicle status characteristics: average battery SOC (statistics from charging pile data), number of vehicles waiting to be charged.

[0085] Historical response characteristics: the actual adjustment amount under similar scenarios in the past 24 hours.

[0086] Model Output: For each future time period t, the model outputs the predicted range of adjustable load power for that time period. As for this period The hard constraints are defined as upper and lower bounds. This range essentially reflects the physical limits of load adjustability under given characteristics.

[0087] Hybrid stochastic / robust optimization handles dual uncertainties: Contributing to renewable energy S typical scenarios generated based on historical data clustering were used. To depict, among which This represents the probability of a scenario.

[0088] For real-time electricity prices Robust modeling using budget uncertainty sets:

[0089] in, It is a predicted value. It is the magnitude of the prediction error. It is a conservatism parameter (budget), which controls the maximum number of time periods for which the electricity price can simultaneously take the worst value.

[0090] Ensemble Model: The final second-stage problem is a min-max problem: given... Below, regarding the worst-case scenario of real-time electricity prices. The goal is to minimize the running and deviation costs in this scenario. Using the duality theory of linear programming, this inner max problem can be transformed into a min problem, thus reconstructing the entire two-phase problem into a large-scale single-phase mixed-integer linear programming model. This model simultaneously considers the stochasticity of the scenario and the robustness of the price.

[0091] Step 3: Multi-mode adaptive switching mechanism based on rolling time domain and state discrimination This step integrates the three models mentioned above into a dynamically running online framework to achieve intelligent switching.

[0092] S3.1 Scrolling Temporal Optimized Execution Engine Clock synchronization: Set a global scheduling clock at fixed time intervals. A scheduling cycle is triggered (e.g., every 15 minutes).

[0093] Scrolling window: at the beginning of each cycle The system obtains the latest status measurement values ​​( , (actual load, etc.), and calls the prediction module to generate a forecast of the future. Time period ( arrive The predicted data.

[0094] Mode identification and model invocation: Based on the logic of S3.2, determine the current operating mode to be adopted, and invoke the corresponding optimization model (S2.1, S2.2 or S2.3) to optimize and solve the future window with the current state as the initial condition.

[0095] Command issuance and execution: Only the current time period from the optimization results is used. Control commands ( , , (etc.), which are then distributed to field equipment through the execution layer.

[0096] State Updates and Rollover: Waiting for the Next Cycle Repeat the above process. The prediction window then scrolls forward.

[0097] S3.2 Embedded Multi-Condition Pattern Intelligent Judgment Logic In each At each time point, the following judgments are executed sequentially: Emergency mode priority criteria: Hard alarm: Directly reads the fault hard signals (such as circuit breaker tripping, protection action) from the equipment monitoring system. If any are detected, immediately set the mode flag to "emergency".

[0098] Soft early warning: Query the equipment health prediction system to obtain the future health forecast of key equipment (main transformer, core diesel engine). Failure probability prediction value within the time period .like

[0099] Then set the mode to "Emergency".

[0100] Market opportunity criteria (judgment only in non-emergency scenarios): Arbitrage potential calculation: Calculate the expected arbitrage index within the future optimization window. :

[0101] in, Weighting (higher weighting for peak periods). Based on cost, The standard deviation of electricity price fluctuations This represents the risk preference coefficient.

[0102] System flexibility status verification: Energy storage status: (e.g., [25%, 85%]).

[0103] Adjustable load potential: From The sum of the adjustment boundaries for the current and future time periods obtained by the model Greater than the threshold.

[0104] Triggering condition: If If the flexibility check passes, then the mode is set to "Market".

[0105] Default mode: If none of the above criteria are met, the mode is set to "Economic".

[0106] S3.3 Smooth Transition Guarantee Strategy To prevent sudden changes in control commands during switching, the following combined strategy is adopted: State inheritance: In any mode, the initial state of the optimization model is strictly equal to the actual measured state at the end of the previous time period, ensuring the physical continuity of the dynamic process.

[0107] Common constraint preservation: The optimization problem of the three models shares the basic physical and safety constraints of all devices (such as SOC limits, generator ramp rates, and grid tie-line power limits). This ensures that the commands solved by different models are physically feasible and that no limit violations occur.

[0108] Transition Period Softening Constraints: When a mode shift is detected (e.g., from economic to market), the first few periods after the corresponding shift point in the optimization model are designated as productive load periods. Add an additional "soft constraint on the rate of power change" to allow it to slowly transition to the optimal value in the new mode, avoiding disruption to the production process.

[0109] Output instruction filtering: Before the instruction is issued, a first-order low-pass filter is applied to the current instruction value and the actual instruction value of the previous cycle. ,in The filter coefficient can be set to a smaller value (e.g., 0.3) during mode switching to ensure a smooth transition.

[0110] Online self-learning and parameter adaptation of the S3.4 model Online updates to the prediction model: Employing online learning algorithms for time series models, such as recursive parameter estimation for the ARIMA model, or online fine-tuning of neural networks, to enable predictions to continuously adapt to the latest data patterns.

[0111] Data-driven model Iteration: Establish a closed-loop data flow. Use actual electricity prices, production status, and corresponding actual load response data as new samples, periodically (e.g., daily) adding them to the training set. The model is incrementally updated or fully retrained to continuously approximate the true response relationship.

[0112] Dynamic parameter tuning: Based on operational performance over a period of time (e.g., whether the emergency mode successfully ensures coverage, and the actual return vs. predicted return of the market mode), key parameters, such as the Wasserstein radius, are dynamically adjusted through Bayesian optimization or trial and error. Robust budgeting Risk tolerance and mode switching threshold These measures enable continuous optimization of system performance.

[0113] Step 4: System Integration, Deployment, and Online Operation All the above modules are integrated into a complete software system and deployed in the mine dispatch center.

[0114] S4.1 Software Architecture: A microservice architecture is adopted. Core services include: data acquisition and preprocessing service, prediction service, optimization engine service (embedded with three model solvers), pattern discrimination service, and instruction distribution service. Services communicate asynchronously through message queues (such as RabbitMQ) to ensure decoupling and reliability.

[0115] S4.2 Hardware and Communication: The scheduling server adopts a high-availability cluster. It communicates with field PLCs and intelligent devices through industrial gateways and mining explosion-proof communication equipment, using protocols such as OPC UA and Modbus TCP.

[0116] S4.3 Human-Computer Interaction: Provides a web-based visual interface to display real-time operating status, mode information, optimization results, key indicators (cost, benefit, reliability), and supports manual intervention mode (forced switching, parameter modification).

[0117] S4.4 Startup and Operation: After the system is powered on, it automatically executes the rolling cycle of S3.1 to achieve uninterrupted automatic optimization scheduling 24 / 7.

[0118] Alternative solutions description The technical solution proposed in this embodiment is a systematic framework in which there are alternative technical paths for multiple links. These alternative solutions can also achieve the purpose of this invention, but may differ in performance, complexity or applicable conditions.

[0119] Alternatives to load classification and modeling include: more refined classification, alternative modeling methods (such as load cluster models), and alternative data-driven algorithms (random forests, LSTM-Attention, Gaussian process regression).

[0120] Alternatives to emergency reliability modeling: different probabilistic distances ( - Divergence, total variation distance), different robust frameworks (robust optimization + CVaR), scene generation method alternatives (historical resampling, fault tree analysis).

[0121] Alternatives to handling uncertainty in market participation: pure stochastic programming, pure discrete bar optimization, and interval optimization.

[0122] Alternatives to the mode switching mechanism include: discriminators based on fuzzy logic or decision trees, switching strategies based on reinforcement learning, and direct solutions for multi-objective optimization.

[0123] Alternative to the overall architecture: layered distributed optimization.

[0124] The following section uses a large-scale open-pit iron mine microgrid demonstration project as an example to illustrate the implementation of the present invention in detail.

[0125] 1. System Configuration Power supply: 5MW rooftop solar power, 3MW wind turbine, 2MW / 8MWh lithium iron phosphate battery energy storage, and 4 1.5MW diesel generators.

[0126] Load: Production load (electric shovel, crusher) maximum 3.5MW, critical load (ventilation, drainage) 1MW, adjustable load (electric mining truck charging pile group) maximum 2MW, basic load 0.5MW.

[0127] Power Grid and Market: Connected to the grid via a 35kV line, with a maximum interactive power of 6MW; connected to the provincial day-ahead spot market.

[0128] 2. Implementation Steps 2.1 Data Foundation and Model Building Data acquisition system: Deploy smart meters and sensors to transmit data back via a 5G private network.

[0129] Digital twin platform: Constructs 3D models to simulate faults such as "main transformer tripping" and "photovoltaic damage", generating 1,500 sets of samples.

[0130] Load characteristic analysis: Based on historical data, the baseline and adjustable range of four types of loads are calibrated.

[0131] Model training: Economic model: Determine cost parameters.

[0132] Emergency model: Set $\epsilon=0.01$, cross-validation yields $\delta=0.08$.

[0133] Market Model: Two years of data were collected to train an XGBoost model to predict the power boundary of charging piles. A two-stage optimization model was constructed, with the budget parameter $\Gamma_{\lambda}=6$ set for the uncertainty set of real-time electricity prices.

[0134] 2.2 Software Deployment and Integration The scheduling core is developed using Python and integrates the Gurobi solver.

[0135] Write the main program for rolling optimization, with a cycle of 15 minutes.

[0136] Deployed on an Intel Xeon server, it communicates with the field control system via the OPC UA protocol.

[0137] 2.3 Execution Test and Results Economic model: Operating on days with no faults and small price differences, the average daily operating cost is reduced by 18% compared to traditional rule-based scheduling.

[0138] Emergency mode: Simulates diesel engine failure, the system switches and ensures 100% power supply to critical loads within 1 minute.

[0139] Market model: On peak electricity price days, the highest net profit per day can reach 6,200 yuan through energy storage arbitrage and load adjustment.

[0140] Switching performance: 27 switching operations were conducted within the quarter without causing sudden power changes or voltage overruns; the process was smooth.

[0141] 2.4 Computational Performance The total time for each 15-minute scheduling cycle is approximately 33 seconds in economic mode, 82 seconds in market mode, and 90 seconds in emergency mode, all of which meet the real-time requirements.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-scenario collaborative optimization scheduling method for microgrids in mining scenarios, characterized in that, include: Step 1: Data preprocessing and basic system modeling tailored to the characteristics of the mine: S1.1 Establish a unified data acquisition platform to synchronously collect power data, load data, environmental and market data at a frequency of no less than 1 minute; adopt the "3σ criterion" in combination with the upper and lower limits of equipment operation to identify and remove outliers; use time series prediction model imputation or decision tree-based method to fill missing data; and use wavelet transform for smoothing filtering of photovoltaic data. S1.2 Load Refinement Classification Modeling Based on Physical Mechanisms and Data-Driven Approach: Decomposing the total load into four independent variables: Productivity load: Establish a mathematical model based on equipment start-up and shutdown status, rated power, and load rate, and set constraints on minimum continuous operation time, minimum downtime, and workload requirements; Critical load protection: Establish a mathematical model of rated power and allowable fluctuations as a hard constraint to ensure power supply continuity; Adjustable load: Establish a flexible mathematical model that includes baseline planning and up / down power adjustment, and set dynamic boundary constraints and energy conservation constraints, wherein the adjustable power boundary is dynamically output by the data-driven model; Basic load: Forecasted using a time series model; S1.3 LSTM neural network combined with production calendar is used for load forecasting, physical-statistical hybrid model combined with numerical weather prediction is used for photovoltaic power output forecasting, power conversion model based on wind speed forecasting is used for wind power output forecasting, and GARCH model or deep probabilistic forecasting model is used for electricity price forecasting. Step Two: Constructing Three Core Optimization Scheduling Models: S2.1 With minimizing the total operating cost within the scheduling cycle as the objective function, a mixed integer linear programming model is established, which includes refined power balance constraints, detailed energy storage model, diesel generator start-stop logic constraints, power grid interaction constraints, and adjustable load constraints, to construct a refined benchmark economic optimal scheduling model; S2.2 Based on the mine digital twin system, an emergency scenario library is constructed. The model is modeled using the joint chance constraint of the sub-Brussels bar based on Wasserstein distance. The probabilistic constraints are equivalently transformed into deterministic linear constraints. With the goal of minimizing emergency fuel costs or maximizing safe time periods, an emergency standby scheduling model based on the chance constraint of the sub-Brussels bar is constructed. S2.3 Establish a two-stage decision-making framework. The first stage determines the day-ahead market bidding volume and generator start-up and shutdown status. The second stage determines the real-time scheduling of internal resources in response to the uncertainties of renewable energy output and real-time electricity prices. A machine learning model is used to model the dynamic feasible region of adjustable loads. A hybrid stochastic / robust optimization is used to handle the dual uncertainties. The two-stage problem is reconstructed into a single-stage mixed integer linear programming model. A collaborative optimization model for virtual power plants participating in the day-ahead spot market is constructed. Step 3: Multi-mode adaptive switching mechanism based on rolling time domain and state discrimination: S3.1 Trigger the scheduling cycle at fixed time intervals, obtain the latest state measurement values ​​and prediction data, call the corresponding optimization model to solve, only use the control instructions for the current time period to be issued and executed, and wait for the next cycle to repeat the above process; S3.2 The emergency mode priority criterion and the market mode opportunity criterion are executed sequentially to determine the current operating mode based on the equipment fault signal, the fault probability prediction value, the expected arbitrage index, and the system flexibility status. S3.3 Smooth transition during execution mode switching is ensured by a combination of strategies including state inheritance, preservation of common constraints, softening of constraints during transition periods, and filtering of output instructions to prevent sudden changes in control instructions during switching. S3.4 The prediction model is updated using an online learning algorithm, a closed-loop data flow is established to iteratively update the data-driven model, and the optimization parameters are dynamically adjusted based on the running performance. Step 4: Integrate the various modules using a microservice architecture and deploy them in the mine scheduling center to achieve uninterrupted automatic optimization scheduling.

2. The method according to claim 1, characterized in that, The dynamic boundary constraints and energy conservation constraints in the adjustable load of S1.2 are as follows: Dynamic boundary constraints: ; Energy conservation constraint: ; in, and Data-driven model The system dynamically outputs data based on market characteristics, production characteristics, time characteristics, vehicle status characteristics, and historical response characteristics.

3. The method according to claim 1, characterized in that, The joint chance constraint modeling of the split-bars based on Wasserstein distance in S2.2 specifically includes: Define the power supply margin function at time t: ， in For emergency dispatch decision variables; Construct joint opportunity constraints: ; Among them, fuzzy set Defined as: , The distance is 1-Wasserstein. Let be the radius.

4. The method according to claim 3, characterized in that, The proposed bibliometric joint chance constraint is transformed into a mixed integer linear programming constraint set by introducing auxiliary variables, thus completely converting the complex probabilistic constraints into deterministic linear and integer constraints.

5. The method according to claim 1, characterized in that, The data-driven dynamic feasible region modeling of adjustable loads described in S2.3 specifically includes: Training machine learning models Input features include: Market characteristics: Day-ahead electricity prices, real-time electricity price forecasts, and their volatility; Production characteristics: shift schedule, planned number of vehicles dispatched, current shift's charged battery capacity, average transport distance; Time characteristics: moment, weekday / holiday, season; Vehicle status characteristics: average battery SOC, number of vehicles waiting to be charged; Historical response characteristics: actual adjustment amounts under similar scenarios in the past 24 hours; The model outputs the prediction range of adjustable load power for that period. As for this period The upper and lower bounds of the hard constraints.

6. The method according to claim 1, characterized in that, The hybrid stochastic / robust optimization approach described in S2.3 specifically includes the following: For renewable energy output, S typical scenarios generated by clustering historical data are used to characterize it; For real-time electricity prices, robust modeling is performed using a budget uncertainty set: ， in, It is a predicted value. It is the magnitude of the prediction error. It is a conservatism parameter, which controls the maximum number of time periods during which the electricity price can simultaneously take the worst value; By utilizing the duality theory of linear programming, the inner-level max problem is transformed into a min problem, and the two-stage problem is reconstructed into a single-stage mixed-integer linear programming model.

7. The method according to claim 1, characterized in that, The emergency mode priority criteria described in S3.2 include: Hard alarm: Directly read the fault hard signal from the equipment monitoring system. If any is found, immediately set the mode flag to "emergency". Soft early warning: Query the equipment health prediction system to obtain information on the future health of critical equipment. Failure probability prediction value within the time period ,like Then set the mode to "Emergency".

8. The method according to claim 1, characterized in that, The market pattern opportunity criteria described in S3.2 include: Calculate the expected arbitrage index within the future optimization window: : ， in, To optimize the expected arbitrage index within the future window, Indicates the current moment. This indicates an optimization of the window length. express Time weight, express The daytime electricity price express The benchmark electricity price at any given time. This is the risk aversion coefficient, used to balance arbitrage profits with the risk of electricity price fluctuations. The standard deviation of the day-ahead electricity price represents the risk of electricity price volatility. Perform system flexibility status verification to determine whether the energy storage status is within a safe range and whether the adjustable load potential is greater than the threshold. like If the flexibility check passes, the mode is set to "Market"; otherwise, the mode is set to "Economy".

9. The method according to claim 1, characterized in that, The smooth transition guarantee for mode switching described in S3.3 includes: State inheritance: In any mode, the initial state of the optimized model is strictly equal to the actual measured state at the end of the previous time period; Common constraints are maintained: the three models share the basic physical and security constraints of all devices; Transition period soft constraint: When it is determined that the mode is about to switch, add an additional soft constraint on the rate of change of power for the productive load; Output command filtering: A first-order low-pass filter is used to smooth the command value, and a smaller filter coefficient is used at the mode switching time.

10. The method according to claim 1, characterized in that, The online self-learning and parameter adaptation of the model described in S3.4 include: Online updates to the prediction model: using recursive parameter estimation of the ARIMA model or online fine-tuning of the neural network; Data-driven model Iteration: Establish a closed-loop data flow, and periodically add actual execution data as new samples to the training set for incremental updates or full retraining; Optimize parameters dynamically: Based on the running performance, dynamically adjust the Wasserstein radius, robust budget, risk tolerance, and mode switching threshold through Bayesian optimization or trial and error.

Images