Multi-time scale energy optimization scheduling method
By employing a multi-timescale energy optimization scheduling method and utilizing technologies such as diffusion models and DC-TSP, automated testing of energy storage systems has been achieved, reducing testing costs, improving testing efficiency and system stability, and solving the problem of low efficiency in manual testing in existing technologies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-04-07
AI Technical Summary
The testing process for energy storage systems is costly, time-consuming, and labor-intensive, and it is not conducive to widespread application. Existing methods rely on manual operation, which is inefficient.
A multi-timescale energy optimization scheduling method is adopted. Tail scenarios are generated through a diffusion model and incorporated into robust optimization. Multi-scale time-series features are extracted using DC-TSP. Combined with Auto-FRL and safety barrier projection, millisecond-level control is achieved. NeuralODE accurately describes energy storage degradation, and dual-loop verification ensures the global consistency of day-ahead, intraday, and real-time solutions.
It has enabled the automation and efficiency of energy storage system testing, reduced testing costs, improved testing efficiency, and ensured the stability and consistency of the system.
Smart Images
Figure BDA0005669673980000061 
Figure BDA0005669673980000086 
Figure FDA0005669673970000021
Abstract
Description
Technical Field
[0001] This invention relates to the field of automated testing technology, specifically a multi-timescale energy optimization scheduling method. Background Technology
[0002] Energy storage has broad application prospects and will create enormous value for my country's economic growth and green energy development in the future. However, as an emerging technology industry, it still faces some challenges at this stage. Improving its technological and economic viability, perfecting its application market mechanisms and pricing systems are key priorities for the future. Most urgently, however, is the need for collaborative efforts from all parties to explore and develop multiple commercially viable markets for the energy storage industry, ensuring its healthy and sustainable development. With the successive introduction of relevant support policies from governments worldwide, investment in the energy storage market is continuously increasing, the industrial chain is constantly improving, business models are becoming increasingly diversified, and application scenarios are rapidly expanding. Domestically, a series of policies have accelerated the vigorous development of the energy storage industry.
[0003] Energy storage systems require a series of processes before shipment, including pack assembly, whole-unit testing, aging and inspection, and final packaging. The common method is to periodically schedule on-site discharge tests, which is costly, time-consuming, and labor-intensive. This hinders the widespread adoption of lithium battery systems.
[0004] Whole-system testing of energy storage systems: Whole-system testing requires testers to be familiar with the testing procedures and the standard data ranges for test items such as temperature, current, voltage, and SOC of the energy storage system. During testing, testers need to manually input test commands using a third-party serial port debugging tool and manually judge whether the tested energy storage system passes each test item based on the data fed back by the test equipment. If the test passes, a qualified label is affixed to the tested energy storage system; if the test fails, the tested energy storage system will be returned to the factory for repair or disassembled and reassembled. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, this invention provides a multi-timescale energy optimization scheduling method. It generates tail scenarios through a diffusion model and incorporates them into robust optimization; it utilizes DC-TSP to extract multi-scale time-series features; it combines Auto-FRL with safety barrier projection to achieve millisecond-level control; it uses GNN to quickly approximate power flow, and NeuralODE to accurately describe energy storage degradation; and it employs dual-loop verification to ensure the global consistency of day-ahead, intraday, and real-time solutions.
[0006] To achieve the above objectives, a multi-timescale energy optimization scheduling method is designed, comprising an energy system, characterized in that the energy system includes:
[0007] Data and Constraint Aggregation Layer: Includes market data, NWP weather data, historical load / PV data, equipment parameters and network topology, equipment rating / safety data, and O&M constraints and compliance data, with unified timestamps, cleaning, and completion; Feature Engineering and State Estimation: Generates holiday / period features, temperature-load interaction data, SOC / temperature estimation, load decomposition, and anomaly detection;
[0008] Generative Scenario Library: Conditional diffusion models generate multivariate scenarios, including electricity price, load, renewable power, weather, and equipment availability.
[0009] Time series predictors include DC-TSP and Koopman aids. DC-TSP generates day-ahead and intraday baseline sequences; Koopman aids provide linearly interpretable evolution of exogenous variables.
[0010] GNN power flow proxy: Graph neural network approximates power flow and node constraints for fast feasibility determination;
[0011] Recently, a robust optimization method was proposed: finding robust solutions to the scenario set within the Wasserstein sphere, imposing chance constraints, and outputting the 24-hour plan and backup curves.
[0012] Intraday rolling consistency correction: Rolling window every 5 to 15 minutes, minimizing cost + deviation from the previous day's solution + soft handover, linked to short-term conditional scenario verification;
[0013] Real-time controller: millisecond / second-level policies; rewards are synthesized using automated preferences, and instructions are projected out through safety barriers / feasibility domains;
[0014] Health and lifespan modeling: Based on a continuous degradation model of temperature / current / SOC trajectory, it gives SOH and cyclic penalty to constrain long-term operation and maintenance;
[0015] Consistency Coordination and Dual-Loop Verification: Scenario Robust Loop: Replaying Difficult Scenarios Across Layers; Consistency Loop: Penalizing Cross-Layer Deviations with Lagrange Multipliers;
[0016] Training and distillation: Distill complex strategies into lightweight models for deployment; offline retraining to absorb new data and scenarios; monitoring and auditing: indicator dashboards, compliance audits, backtracking and review, generating reports and alerts;
[0017] The optimized scheduling method flow is as follows:
[0018] (1) Data input, including market data, NWP weather, historical load / photovoltaic, equipment parameters and network topology, equipment rating / safety data, operation and maintenance restrictions and compliance data;
[0019] (2) Core computation, including deduplication, alignment, interpolation, anomaly detection, clock unification and time zone normalization, and constraint archiving into structured sets and topology graphs;
[0020] (3) Output structure, including timing base table, topology, and device constraints.
[0021] The feature engineering and state estimation process is as follows:
[0022] S11, Input: T, Where T is the timing base table. For network topology, For constraint set;
[0023] S12, Calculate:
[0024] Calendar / holiday / time period coding, temperature-load interaction terms, lag and sliding window statistics;
[0025] SOC / temperature estimation yields the state vector x t ;
[0026] Load decomposition and outlier correction;
[0027] S13, Output feature matrix X, state vector stream x t .
[0028] The process of the generative scenario library is as follows:
[0029] S21, Input: Feature matrix X, State vector stream x t With conditional vector c;
[0030] S22, Training loss calculation: in, Here are the parameters of the denoising network; y represents a clean sample; ∈ represents Gaussian noise; and t represents the diffusion time step. For the noise prediction subnetwork; α t , σ t The time-dependent signal / noise figure; |·|2 is the L2 norm;
[0031] S23, Sampling: DDIM / CFG acceleration, importance resampling amplifies the tail;
[0032] S24, Output: Scenario set Ω. Each scenario contains a combined trajectory of electricity price, load, renewable energy, and weather.
[0033] The process of the time series predictor is as follows:
[0034] S31, Input: Feature matrix X, State vector stream x t ;
[0035] S32, DC-TSP model calculation: in, Let d be the void ratio of the kth group. k The input slice below; b (k) For convolution kernel and bias; K is the number of scale groups; σ(·) is the non-linear activation; [·|·] is the feature concatenation; W o b o For output layer parameters; For τ-step prediction;
[0036] S33, Koopman Auxiliary: Constructing approximate linear evolution for exogenous variables to enhance interpretability;
[0037] S34, Output: Baseline Forecast, including electricity price forecast Load forecasting Renewable forecast
[0038] The process of the GNN current flow proxy is as follows:
[0039] S41, Input network topology Injected power;
[0040] S42, Message Passing Formula: in, Let N(v) be the latent feature of node v in layer l; N(v) be the set of neighbors of node v; e uv Let be the feature of edge (u,v); ψ(·) is the message function; For update functions;
[0041] S43, Output: Approximate node voltage, line current, and network loss, forming a fast evaluation feasible region.
[0042] The process of optimizing the day-ahead blue bar is as follows:
[0043] S51, Input scenario set Ω, electricity price forecast Load forecasting feasible region constraint set
[0044] S52, Calculate the objective function: Among them, s t For SOC; π t Electricity price; Δp t Energy storage ramp-up; g t For the purchase and sale of electrical power; u D The current-day decision sequence is given; Q is the distribution to be determined. Based on empirical distribution A Wasserstein sphere centered at π with radius δ; t(ω) represents the electricity price at time t under scenario ω; λ c Φ(·) represents the ramp-up cost weight; ρ represents the SOC penalty;
[0045] S53, Opportunity Constraint: P ω～Q (g j (u D ,ω)≤0)≥1-α,; Among them, g j (·) represents the j-th operational safety constraint; α represents the upper limit of the allowed probability of default;
[0046] S54, Output Reference Curve With reserve capacity; where reserve capacity is determined by the upward / downward adjustment curve and constraint set. Specify the callable boundaries; For optimal work efficiency; The optimal reactive power; Optimal power purchase and sale; Optimal SOC.
[0047] The process for intraday rolling consistency correction is as follows:
[0048] S61, Enter the latest electricity price forecast Short-term conditional scenario Ω short Day-ahead benchmark
[0049] S62, Optimized Calculation: Where η is the soft handover coefficient; γ is the penalty coefficient for day-ahead baseline consistency; u I For intraday decision-making; C t (·) represents the rolling cost; |·|1, |·|2 are the norms;
[0050] S63, Output rolling correction solution u I And generate the startup value for the next cycle.
[0051] The process of the real-time controller is as follows:
[0052] S71, Input the SOC value s of the current state. t Feasible region F;
[0053] S72, Strategy Output: u R,t =Proj F (π θ (s t )); where π θ (s t ) represents the policy in state s t The original action output; F is the real-time feasible region; Proj F (·) is used to project the action into F, ensuring it does not exceed the bounds;
[0054] S73, Strategy Optimization: Where, π θ For strategy; π θ- This is the old strategy; This is the advantage function estimate; clip(·) is the truncation operator that restricts the policy step size; KL(·|·) is the KL divergence; β is the regularity coefficient;
[0055] S74, Reward Automation: Frequency / Voltage Deviation Penalty, Soft Penalty for Out-of-Limit, Smoothing of Ramp-up, Energy Cost, Lifetime Cost.
[0056] The process for health and lifespan modeling is as follows:
[0057] S81, Evolution and Observation: Where z(t) represents the degenerate hidden state; u(t) represents the charge / discharge power / current; T(t) represents the temperature; and f ξ (·) represents the evolutionary network; g ξ (·) represents the observation / mapping network; SOH(t) represents the health status;
[0058] S82 outputs SOH and lifetime cost as a penalty or constraint.
[0059] The aforementioned consistency coordination and dual-loop verification include,
[0060] Multiplier Update:
[0061] Consistency penalty weight adaptive:
[0062] Compared with existing technologies, this invention provides a multi-timescale energy optimization scheduling method. It generates tail scenarios through a diffusion model and incorporates them into robust optimization; it uses DC-TSP to extract multi-scale time-series features; it combines Auto-FRL with safety barrier projection to achieve millisecond-level control; it uses GNN to quickly approximate power flow; and it uses NeuralODE to accurately describe energy storage degradation. Dual-loop verification ensures the global consistency of day-ahead, intraday, and real-time solutions. Detailed Implementation
[0063] A multi-timescale energy optimization scheduling method includes an energy system, which comprises:
[0064] Data and Constraint Aggregation Layer: Includes market data, NWP weather data, historical load / PV data, equipment parameters and network topology, equipment rating / safety data, and O&M constraints and compliance data, with unified timestamps, cleaning, and completion; Feature Engineering and State Estimation: Generates holiday / period features, temperature-load interaction data, SOC / temperature estimation, load decomposition, and anomaly detection;
[0065] Generative Scenario Library: Conditional diffusion models generate multivariate scenarios, including electricity price, load, renewable power, weather, and equipment availability.
[0066] Time series predictors include DC-TSP and Koopman aids. DC-TSP generates day-ahead and intraday baseline sequences; Koopman aids provide linearly interpretable evolution of exogenous variables (such as irradiance and temperature).
[0067] GNN Power Flow Proxy (PF Proxy): A graph neural network that approximates power flow and node constraints (voltage, line current, loss) for rapid feasibility assessment;
[0068] Recently, a robust optimization method was proposed: finding robust solutions to the scenario set within the Wasserstein sphere, imposing chance constraints, and outputting the 24-hour plan and backup curves.
[0069] Intraday rolling consistency correction: Rolling window every 5 to 15 minutes, minimizing cost + deviation from the previous day's solution + soft handover, linked to short-term conditional scenario verification;
[0070] Real-time controller: millisecond / second-level policies; rewards are synthesized using automated preferences, and instructions are projected out through safety barriers / feasibility domains;
[0071] Health and lifespan modeling: Based on a continuous degradation model of temperature / current / SOC trajectory, it gives SOH and cyclic penalty to constrain long-term operation and maintenance;
[0072] Consistency Coordination and Dual-Loop Verification: Scenario Robust Loop: Replaying Difficult Scenarios Across Layers; Consistency Loop: Penalizing Cross-Layer Deviations with Lagrange Multipliers;
[0073] Training and Distillation: Distill complex strategies into lightweight models for deployment; offline retraining to absorb new data and scenarios; Monitoring and Auditing: Indicator dashboards, compliance audits, backtracking and review (plan-execution reconciliation), generating reports and alerts; Optimized scheduling methods and processes are as follows:
[0074] (1) Data input, including market data, NWP weather (irradiance, temperature, speed), historical load / photovoltaic, equipment parameters and network topology, equipment rating / safety data, operation and maintenance restrictions and compliance data;
[0075] (2) Core computation, including deduplication, alignment, interpolation, anomaly detection (box line or based on reconstruction error), clock unification and time zone normalization, constraint archiving into structured sets and topology graphs;
[0076] (3) Output structure, including timing base table, topology, and device constraints. As shown in Table 1.
[0077] Table 1
[0078]
[0079] The process of feature engineering and state estimation is as follows:
[0080] S11, Input: T, Where T is the timing base table. For network topology, For constraint set;
[0081] S12, Calculate:
[0082] Calendar / holiday / time period coding, temperature-load interaction terms, lag and sliding window statistics;
[0083] SOC / temperature estimation (Kalman or particle filter) yields the state vector x t ;
[0084] Load decomposition (e.g., STL) and outlier correction;
[0085] S13, Output feature matrix X, state vector stream x t As shown in Table 2.
[0086] Table 2
[0087] name structure Dimension Characteristic matrix X float32 T×d <![CDATA[State vector flow x t > dict Every moment
[0088] The process of generative scenario library is as follows:
[0089] S21, Input: Feature matrix X, State vector stream x t With condition c (such as "initial day-ahead curve, short-term weather, equipment status");
[0090] S22, Training loss calculation: in, denoising network parameters; y represents clean samples (such as historical trajectory slices); ∈ represents Gaussian noise; t represents the diffusion time step; For the noise prediction subnetwork; α t , σ t The time-dependent signal / noise figure; |·|2 is the L2 norm;
[0091] S23, Sampling: DDIM / CFG acceleration, importance resampling amplifies the tail;
[0092] S24, Output: Scenario set Ω. Each scenario contains a combined trajectory of electricity price, load, renewable energy, and weather.
[0093] The process of the timing predictor (DC-TSP / Koopman) is as follows:
[0094] S31, Input: Feature matrix X, State vector stream x t ;
[0095] S32, DC-TSP model calculation: b o ,in, Let d be the void ratio of the kth group. k The input slice below; b (k) For convolution kernel and bias; K is the number of scale groups; σ(·) is the non-linear activation; [·|·] is the feature concatenation; W o b o For output layer parameters; For τ-step prediction;
[0096] S33, Koopman Auxiliary: Constructing approximate linear evolution for exogenous variables to enhance interpretability;
[0097] S34, Output: Baseline Forecast, including electricity price forecast Load forecasting Renewable forecast
[0098] The process for becoming a GNN fashion distributor is as follows:
[0099] S41, Input Topology Injected power (from S3 / S4 output scenario / prediction);
[0100] S42, Message Passing Formula: in, Let N(v) be the latent feature of node v in layer l; N(v) be the set of neighbors of node v; e uv Let be the characteristics (impedance, rating, length, etc.) of the edge (u,v); ψ(·) is the message function; For update functions;
[0101] S43, Output: Approximate node voltage, line current, and network loss, forming a fast evaluation feasible region.
[0102] The current optimization process for the Bluebar is as follows:
[0103] S51, Input scenario set Ω, electricity price forecast Load forecasting feasible region constraint set
[0104] S52, Calculate the objective function: Among them, s t For SOC; π t Electricity price; Δp tEnergy storage ramp-up; g t For the purchase and sale of electrical power; u D The current-day decision sequence is given; Q is the distribution to be determined. Based on empirical distribution A Wasserstein sphere centered at π with radius δ; t (ω) represents the electricity price at time t under scenario ω; λ c Φ(·) represents the ramp-up cost weight; ρ represents the SOC penalty;
[0105] S53, Opportunity Constraint: P ω～Q (g j (u D ,ω)≤0)≥1-α,; Among them, g j (·) represents the j-th operational safety constraint (such as voltage, line current, power balance); α represents the upper limit of the allowable probability of default (risk level);
[0106] S54, Output Reference Curve With reserve capacity; where reserve capacity is determined by the upward / downward adjustment curve and constraint set. Specify the callable boundaries; For optimal work efficiency; The optimal reactive power; Optimal power purchase and sale; Optimal SOC.
[0107] The process for intraday rolling consistency correction is as follows:
[0108] S61, Enter the latest electricity price forecast Short-term conditional scenario Ω short Day-ahead benchmark
[0109] S62, Optimized Calculation: Where η is the soft handover (smoothing) coefficient; γ is the penalty coefficient for day-ahead baseline consistency; u I For intraday decision-making; C t (·) represents the rolling cost; |·|1, |·|2 are the norms;
[0110] S63, Output rolling correction solution u I And generate the startup value for the next cycle.
[0111] The real-time controller process is as follows:
[0112] S71, Input the SOC value s of the current state. t Feasible region F;
[0113] S72, Strategy Output: u R,t =ProjF (π θ (s t )); where π θ (s t ) represents the policy in state s t The original action output; F is the real-time feasible region (GNN feasible region approximation + hardware limiting); Proj F (·) is used to project the action into F, ensuring it does not exceed the bounds;
[0114] S73, Strategy Optimization: Where, π θ For strategy; π θ- This is the old strategy; This is the advantage function estimate; clip(·) is the truncation operator that restricts the policy step size; KL(·|·) is the KL divergence; β is the regularity coefficient;
[0115] S74, Reward Automation: Frequency / Voltage Deviation Penalty, Soft Penalty for Out-of-Limit, Smoothing of Ramp-up, Energy Cost, Lifetime Cost.
[0116] The process of health and lifespan modeling is as follows:
[0117] S81, Evolution and Observation: Where z(t) represents the degenerate hidden state; u(t) represents the charge / discharge power / current; T(t) represents the temperature; and f ξ (·) represents the evolutionary network; g ξ (·) represents the observation / mapping network; SOH(t) represents the health status;
[0118] S82 outputs SOH and lifetime cost as a penalty or constraint.
[0119] Consistency coordination and double-loop verification include,
[0120] Multiplier Update:
[0121] Consistency penalty weight adaptive:
[0122] General sets and timelines:
[0123] Data Structures and Objects: T is the time series base table; For network topology, For the set of equipment / safety constraints.
[0124] State and decision variables: x t The system status includes SOC, temperature, and equipment availability; X∈R T×d p is the characteristic matrix; t Contributing to energy storage; qt For no effect; g t For purchasing and selling electrical power; s t For SOC; Δp t This refers to the amount of uphill climb.
[0125] Probability objects and scenarios: α represents the opportunity-constrained risk level.
[0126] Ω,ω (i) For a set of scenarios / single scenario; π t ,d t ,r t Electricity price / load / renewable energy; d k DC-TSP hidden state / void ratio; g t ,p t ,q t ,s t For purchasing and selling electricity / active power / reactive power / SOC; u R,t For real-time decision-making.
[0127] Example:
[0128] System composition:
[0129] (1) Peak load 12MW, average daily 8MW; rooftop photovoltaic 6MWp (PR≈0.8); energy storage 10MW / 20MWh, SOC window 10–90%.
[0130] (2) There are 22 distribution network nodes, 10kV level, and the thermal limit of the key feeder is 400A.
[0131] (3) The peak-to-valley ratio of electricity prices is about 3.2 (high volatility).
[0132] Key settings:
[0133] (1) There are N=1000 scenarios (diffusion generation + tail amplification) at present, with Wasserstein radius δ=0.12 and risk level α=0.05.
[0134] (2) The intraday rolling interval is 10 minutes, the soft switching weight η = 0.03, and the consistency weight γ = 5.
[0135] (3) Real-time control cycle 500ms; Barrier projection includes voltage 0.95–1.05pu, line current ≤0.95 limit, and SOC 10–90%.
[0136] Key findings (compared to the "Previous Day Only + Simple Rules" baseline):
[0137] (1) Electricity purchase cost: The robust + rolling + real-time solution is lower than the baseline (mainly due to the robust arbitrage between peak discharge and valley charging; intraday rolling reduces mismatch).
[0138] (2) The rate of out-of-bounds (voltage / line current / climbing touch constraint) has decreased (GNN fast agent + projection control is effective).
[0139] (3) End weather days (high temperature + rapid changes in cloud clusters): the curtailment rate decreases (the rolling conditions provide a short-term downtrend scenario, and the SOC is shifted in advance).
[0140] Verifiability:
[0141] (1) Randomly select 7 days to replay the same scenario and price, and repeatedly optimize and control the simulation, with the index fluctuation <±2%.
[0142] (2) Sensitivity test for Dro radius δ: When δ∈[0.08,0.15], the cost and the rate of exceeding the limit are in a convex trade-off, and δ≈0.12 is recommended.
Claims
1. A multi-timescale energy optimization scheduling method, comprising an energy system, characterized in that, The energy system includes: Data and Constraint Aggregation Layer: Includes market data, NWP weather data, historical load / PV data, equipment parameters and network topology, equipment rating / safety data, and O&M constraints and compliance data, with unified timestamps, cleaning, and completion; Feature Engineering and State Estimation: Generates holiday / period features, temperature-load interaction data, SOC / temperature estimation, load decomposition, and anomaly detection; Generative Scenario Library: Conditional diffusion models generate multivariate scenarios, including electricity price, load, renewable power, weather, and equipment availability. Time series predictors include DC-TSP and Koopman aids. DC-TSP generates day-ahead and intraday baseline sequences; Koopman aids provide linearly interpretable evolution of exogenous variables. GNN power flow proxy: Graph neural network approximates power flow and node constraints for fast feasibility determination; Recently, a robust optimization method was proposed: finding robust solutions to the scenario set within the Wasserstein sphere, imposing chance constraints, and outputting the 24-hour plan and backup curves. Intraday rolling consistency correction: Rolling window every 5 to 15 minutes, minimizing cost + deviation from the previous day's solution + soft handover, linked to short-term conditional scenario verification; Real-time controller: millisecond / second-level policies; rewards are synthesized using automated preferences, and instructions are projected out through safety barriers / feasibility domains; Health and lifespan modeling: Based on a continuous degradation model of temperature / current / SOC trajectory, it gives SOH and cyclic penalty to constrain long-term operation and maintenance; Consistency Coordination and Dual-Loop Verification: Scenario Robust Loop: Replaying Difficult Scenarios Across Layers; Consistency Loop: Penalizing Cross-Layer Deviations with Lagrange Multipliers; Training and distillation: Distill complex policies into lightweight models for deployment; offline retraining to absorb new data and scenarios; Monitoring and Auditing: Indicator dashboards, compliance audits, backtesting and review, generating reports and alerts; The optimized scheduling method flow is as follows: (1) Data input, including market data, NWP weather, historical load / photovoltaic, equipment parameters and network topology, equipment rating / safety data, operation and maintenance restrictions and compliance data; (2) Core computation, including deduplication, alignment, interpolation, anomaly detection, clock unification and time zone normalization, and constraint archiving into structured sets and topology graphs; (3) Output structure, including timing base table, topology, and device constraints.
2. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The feature engineering and state estimation process is as follows: S11, Input: T, Where T is the timing base table. For network topology, For constraint set; S12, Calculate: Calendar / holiday / time period coding, temperature-load interaction terms, lag and sliding window statistics; SOC / temperature estimation yields the state vector x t ; Load decomposition and outlier correction; S13, Output feature matrix X, state vector stream x t .
3. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process of the generative scenario library is as follows: S21, Input: Feature matrix X, State vector stream x t With conditional vector c; S22, Training loss calculation: in, Here are the parameters of the denoising network; y represents a clean sample; ∈ represents Gaussian noise; and t represents the diffusion time step. For the noise prediction subnetwork; α t , σ t The time-dependent signal / noise figure; |·|2 is the L2 norm; S23, Sampling: DDIM / CFG acceleration, importance resampling amplifies the tail; S24, Output: Scenario set Ω. Each scenario contains a combined trajectory of electricity price, load, renewable energy, and weather.
4. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process of the time series predictor is as follows: S31, Input: Feature matrix X, State vector stream x t ; S32, DC-TSP model calculation: in, Let d be the void ratio of the kth group. k The input slice below; b (k) For convolution kernel and bias; K is the number of scale groups; σ(·) is the non-linear activation; [·|·] is the feature concatenation; W o b o For output layer parameters; For τ-step prediction; S33, Koopman Auxiliary: Constructing approximate linear evolution for exogenous variables to enhance interpretability; S34, Output: Baseline Forecast, including electricity price forecast Load forecasting Renewable forecast 5. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process of the GNN current flow proxy is as follows: S41, Input network topology Injected power; S42, Message Passing Formula: in, Let N(v) be the latent feature of node v in layer l; N(v) be the set of neighbors of node v; e uv Let be the feature of edge (u,v); ψ(·) is the message function; For update functions; S43, Output: Approximate node voltage, line current, and network loss, forming a fast evaluation feasible region.
6. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process of optimizing the day-ahead blue bar is as follows: S51, Input scenario set Ω, electricity price forecast Load forecasting feasible region constraint set S52, Calculate the objective function: Among them, s t For SOC; π t Electricity price; Δp t Energy storage ramp-up; g t For the purchase and sale of electrical power; u D The current-day decision sequence is given; Q is the distribution to be determined. Based on empirical distribution A Wasserstein sphere centered at π with radius δ; t (ω) represents the electricity price at time t under scenario ω; λ c Φ(·) represents the ramp-up cost weight; ρ represents the SOC penalty; S53, Opportunity Constraint: Among them, g j (·) represents the j-th operational safety constraint; α represents the upper limit of the allowed probability of default; S54, Output Reference Curve With reserve capacity; where reserve capacity is determined by the upward / downward adjustment curve and constraint set. Specify the callable boundaries; For optimal work efficiency; The optimal reactive power; Optimal power purchase and sale; Optimal SOC.
7. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process for intraday rolling consistency correction is as follows: S61, Enter the latest electricity price forecast Short-term conditional scenario Ω short Day-ahead benchmark S62, Optimized Calculation: Where η is the soft handover (smoothing) coefficient; γ is the penalty coefficient for day-ahead baseline consistency; u I For intraday decision-making; C t (·) represents the rolling cost; |·|1, |·|2 are the norms; S63, Output rolling correction solution u I And generate the startup value for the next cycle.
8. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process of the real-time controller is as follows: S71, Input the SOC value s of the current state. t Feasible region F; S72, Strategy Output: u R,t =Proj F (π θ (s t )); where π θ (s t ) represents the policy in state s t The original action output; F is the real-time feasible region; Proj F (·) is used to project the action into F, ensuring it does not exceed the bounds; S73, Strategy Optimization: Where, π θ For strategy; π θ- This is the old strategy; This is the advantage function estimate; clip(·) is the truncation operator that restricts the policy step size; KL(·|·) is the KL divergence; β is the regularity coefficient; S74, Reward Automation: Frequency / Voltage Deviation Penalty, Soft Penalty for Out-of-Limit, Smoothing of Ramp-up, Energy Cost, Lifetime Cost.
9. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The process for health and lifespan modeling is as follows: S81, Evolution and Observation: Where z(t) represents the degenerate hidden state; u(t) represents the charge / discharge power / current; T(t) represents the temperature; and f ξ (·) represents the evolutionary network; g ξ (·) represents the observation / mapping network; SOH(t) represents the health status; S82 outputs SOH and lifetime cost as a penalty or constraint.
10. The multi-timescale energy optimization scheduling method according to claim 1, characterized in that: The aforementioned consistency coordination and dual-loop verification include, Multiplier Update: Consistency penalty weight adaptive: