New energy power system reliability evaluation method based on environment variables

Abstract

The invention provides a new energy power system reliability evaluation method based on environment variables, and relates to the technical field of power systems. According to the method, a non-parametric self-adaptive quantile regression forest model is adopted to perform modeling on wind speed and sunlight intensity, and online incremental updating is triggered based on Kolmogorov-Smirnov statistics; setting a cross-scale state coupling interface layer in the double-layer nested reliability evaluation architecture, defining a shared state vector including an energy storage charge state, an equipment aging index and an available reserve capacity, and applying Lyapunov stability constraint; a period economic value dynamic pricing module based on Q-learning is integrated in a short-term evaluation layer, and a real-time electricity price, a user interruption contract and a load elastic coefficient are used as state spaces to correct a load shedding strategy. According to the scheme, through adaptive modeling of environment variable distribution and cross-scale state stable transmission, the problem of evaluation result lag is effectively solved, and the real-time performance and economical efficiency of reliability evaluation are improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and specifically to a reliability assessment method for new energy power systems based on environmental variables. Background Technology

[0002] As a core component of modern energy structure transformation, the large-scale integration of new energy power systems is of great strategic significance for the safe and stable operation of the power grid. With the continuous growth of installed capacity from renewable energy sources such as wind and solar power, power system reliability assessment has become a crucial link in ensuring power grid planning, dispatching, and operation, and is widely used in regional power grid dispatch centers, new energy power plant monitoring platforms, and power market trading platforms. Existing technologies typically employ parametric probabilistic models to statistically model wind speed and solar radiation intensity, constructing a single-time-scale reliability assessment process. This involves training a fixed-structure prediction model using historical operating data, combining this with Monte Carlo simulation to generate new energy output time-series samples, and finally outputting system reliability indicators. This approach can complete basic assessment tasks under static environmental conditions.

[0003] However, in existing technologies, the dynamic changes in the distribution of environmental variables and the inconsistency in state transfer across time scales make it difficult for the evaluation results to reflect the actual operating status of the system in real time. Summary of the Invention

[0004] This invention provides a reliability assessment method for new energy power systems based on environmental variables, which can solve the problem of lagging assessment results caused by the inconsistency between the dynamic changes in the distribution of environmental variables and the state transmission across time scales.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a reliability assessment method for new energy power systems based on environmental variables. The method includes acquiring historical and predicted environmental variable data, power system topology and operational information, generating time-series samples of new energy output, and constructing a two-layer nested reliability assessment architecture. A non-parametric adaptive quantile regression forest model is used to model wind speed and solar radiation intensity, and online incremental updates are triggered based on Kolmogorov-Smirnov statistics. A cross-scale state coupling interface layer is set in the two-layer nested reliability assessment architecture. This interface layer defines a shared state vector containing energy storage state of charge, equipment aging index, and available reserve capacity, and applies Lyapunov stability constraints. A time-based economic value dynamic pricing module based on Q-learning is integrated into the short-term assessment layer, using real-time electricity prices, user interruption contracts, and load elasticity coefficients as the state space to correct load shedding strategies.

[0007] In one alternative embodiment, the shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, the equipment aging index, and the available reserve capacity, and the shared state vector output by the upper-layer model is used as the initial boundary condition for the lower-layer evaluation process.

[0008] In one alternative embodiment, the energy storage state of charge, equipment aging index, and available reserve capacity in the shared state vector are dynamically coupled to external subsystems through a bidirectional power flow interface, an equipment health monitoring system, and a market clearing and demand response platform, respectively.

[0009] In one alternative embodiment, the two-layer nested reliability assessment architecture includes an upper-layer long-term power transmission system reliability assessment model that operates on an hourly or daily timescale and a lower-layer short-term instantaneous probabilistic reliability assessment process that executes on a minute or second timescale.

[0010] In one alternative embodiment, the shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, the device aging index, and the available reserve capacity, and is passed between the upper and lower layers of the two-layer nested architecture through state transition equations.

[0011] In one optional embodiment, the cross-scale state coupling interface layer is configured with a state transition error bound constraint module based on the Lyapunov function, which is used to mathematically constrain the deviation between the upper layer output state and the lower layer rolling optimization initial state, and to trigger upper layer model parameter recalibration or lower layer initial condition correction when the state transition error exceeds the stability domain.

[0012] In one alternative embodiment, the online incremental update mechanism triggers local parameter retraining or the addition of decision tree nodes to the quantile regression forest model when the Kolmogorov-Smirnov statistic exceeds a preset threshold.

[0013] In one alternative embodiment, the online incremental update mechanism employs a sliding time window structure and assigns different confidence weights to historical and current environmental variable data based on a weighted empirical distribution function. The online incremental update deploys the calculation of Kolmogorov-Smirnov statistics and incremental training tasks on a local processor close to the data source, and uploads the model parameter increments or key statistical summaries to the central evaluation server.

[0014] Compared with the prior art, the beneficial effects of the present invention are:

[0015] 1) By modeling wind speed and solar radiation intensity using a non-parametric adaptive quantile regression forest model, an assumption-free fit to the probability distribution of environmental variables is achieved, avoiding the dependence of parametric models on the prior distribution form; online incremental updates triggered by the Kolmogorov-Smirnov statistic can respond promptly to the drift of environmental variable distribution and maintain the model's prediction accuracy.

[0016] 2) In the two-layer nested reliability assessment architecture, a cross-scale state coupling interface layer is set up to define a shared state vector that includes the energy storage charge state, equipment aging index and available backup capacity, which ensures the state consistency of assessment units at different time scales.

[0017] 3) Applying Lyapunov stability constraints effectively suppresses the accumulation of dynamic deviations during state transfer; integrating a time-based economic value dynamic pricing module based on Q-learning into the short-term evaluation layer uses real-time electricity prices, user interruption contracts, and load elasticity coefficients as the state space to correct load shedding strategies, achieving synergistic optimization of technical security and economic value.

[0018] 4) Through the organic combination of the above technical features, the present invention significantly improves the adaptability of the reliability assessment of new energy power systems to dynamic environmental changes, ensures the stability of state transmission across time scales and the accuracy of economic decision-making, thereby providing reliable technical support for real-time grid dispatch. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit the invention.

[0020] Example 1:

[0021] Traditional reliability assessment methods for new energy power systems generally rely on static probability models or parametric distribution assumptions (such as wind speed following a Weibull distribution and solar radiation intensity following a Beta distribution), making it difficult to characterize the non-stationarity and long-term drift characteristics of meteorological variables under climate change, frequent extreme weather events, and local microclimate disturbances. Furthermore, existing assessment frameworks often employ single-timescale modeling (e.g., conducting N-1 security analysis only at the hourly level), lacking a dynamic coupling mechanism between long-term trend evolution (e.g., the cumulative effect of equipment aging and the seasonal decline of energy storage SOC) and short-term transient risks (e.g., transient voltage instability caused by second-level power fluctuations). This results in state discontinuities, boundary mismatches, and feedback lags in the assessment results between different levels. In addition, key control strategies such as load shedding are typically based on fixed thresholds or empirical rules, failing to incorporate real-time electricity price signals, user interruption contract constraints, and load elasticity response capabilities into the closed-loop optimization process. This leads to a disconnect between reliability assessment and economic operation, making it difficult to synergistically improve overall system operating efficiency and user satisfaction.

[0022] Against this backdrop, this embodiment provides a reliability assessment method for new energy power systems based on environmental variables. This method includes acquiring historical and predicted environmental variable data, power system topology and operational information, generating time-series samples of new energy output, and constructing a two-layer nested reliability assessment architecture. This architecture achieves dynamic adaptation, multi-granularity collaboration, and techno-economic joint optimization for system reliability assessment under high-proportion new energy access through a four-mechanism approach: non-parametric modeling, drift-driven updates, cross-scale state coupling, and economic value reinforcement feedback.

[0023] Step 1: A non-parametric adaptive quantile regression forest model is used to model wind speed and solar radiation intensity, and online incremental updates are triggered based on the Kolmogorov-Smirnov statistic.

[0024] Among them, "non-parametric adaptive quantile regression forest" refers to an ensemble learning structure composed of multiple decision trees. The leaf nodes of each tree do not output the mean or single-point prediction, but instead store the empirical conditional quantile distribution of the training samples at that node (e.g., the 5th, 25th, 50th, 75th, and 95th quantiles), and perform end-to-end optimization through a weighted quantile loss function (such as pinball loss). Its "non-parametric" nature is reflected in the fact that it does not require preset probability density function forms for wind speed or solar radiation intensity (e.g., it does not assume that it follows a Weibull, Lognormal, or Gaussian mixture distribution), and relies solely on historical observation data to drive modeling. Its "adaptive" nature is reflected in the fact that the model structure can be dynamically expanded with the addition of new data. When a distribution shift is detected, local node splitting, subtree pruning, or the addition of weak learners are supported, rather than full model reconstruction. The model input consists of multidimensional environmental covariates, including but not limited to: historical wind speed sequences (sliding window of the previous 6–24 hours), concurrent temperature, pressure gradient, terrain roughness index, satellite cloud image grayscale texture features, and numerical weather prediction (NWP) corrected residuals. The output is a quantile interval prediction (e.g., P10–P90 bandwidth) of wind speed and solar radiation intensity for the next 1–72 hours, supporting the quantification of uncertainty in new energy output. As an optional implementation, the quantile regression forest can be replaced by a variational Bayesian quantile regression network based on kernel density estimation (KDE), or a time-series quantile prediction model (Quantile Transformer) using the Transformer architecture. Both have nonparametric modeling capabilities and are compatible with incremental learning interfaces.

[0025] The "Kolmogorov-Smirnov statistic" refers to a nonparametric statistic used to test the difference in empirical distribution functions between two groups of one-dimensional samples, defined as KS = supx |F n (x)-F m (x)|, where F n (·) and F m (·) represents the empirical cumulative distribution function (ECDF) of n recently collected environmental variable samples within the sliding time window and m historical samples within the baseline window, respectively; in this embodiment, the KS statistic is calculated separately for the two independent dimensions of wind speed and solar radiation intensity, and a differentiation threshold is set (e.g., wind speed KS). 0.01 =0.12, solar radiation intensity (KS) 0.01 =0.09), to accommodate the inherent differences in the fluctuation characteristics of the two; when the KS value of either dimension exceeds the corresponding threshold, it is determined that the environmental variable has undergone a significant distributional shift, triggering the incremental update process. As an optional implementation, the KS statistic can be replaced by Wasserstein distance or Maximum Mean Discrepancy (MMD), both of which are equally applicable to nonparametric distribution comparisons and are more sensitive to tail changes.

[0026] Among them, "online incremental update" refers to the lightweight iterative adjustment process of model parameters or structure, which does not rely on retraining with all historical data; specifically, it includes two modes: (1) Local parameter retraining: freeze the forest trunk structure, and only perform SGD optimization on the quantile estimators of several decision tree tree nodes associated with the geographical area involved in KS drift (such as the meteorological grid where a wind farm or photovoltaic power station is located); (2) Add decision tree nodes: insert 1-3 new trees on the basis of the original forest, and limit their training data to high confidence samples within the drift window (screened by weighted ECDF). The output of the new trees is weighted and fused with the original forest to form the updated quantile prediction. This mechanism is deployed on the edge-side smart gateway or the station-side edge computing unit (such as Huawei Atlas 500, NVIDIA Jetson AGX Orin), and only uploads the model parameter increment (Δθ) or key statistical summary (such as the mean and variance of the offset of each quantile node) to the central evaluation server, which significantly reduces the communication bandwidth occupation and the central computing load. As an optional implementation, incremental updates can be combined with a sliding time window structure, wherein the historical baseline window length is set to 7 days (168 hours), the current window length is set to 24 hours, and an exponentially decaying weight w is applied to samples within the window. t =α t (α = 0.995) gives higher confidence to recent data; this weighted ECDF can be further replaced by a dynamic distribution fitter based on online kernel density estimation.

[0027] Step 2: In the two-layer nested reliability assessment architecture, a cross-scale state coupling interface layer is set up. This interface layer defines a shared state vector that includes the energy storage state of charge, equipment aging index and available backup capacity, and applies Lyapunov stability constraints.

[0028] The "dual-layer nested reliability assessment architecture" is a spatiotemporally decoupled but state-coupled hierarchical assessment paradigm: the upper layer is a long-term power generation and transmission system reliability assessment model, operating on hourly or daily time scales (typically with a step size of 1 hour), focusing on the cumulative impact of planned unit maintenance, transmission corridor N-1 verification, annual load growth trends, and energy storage charging and discharging scheduling strategies on system adequacy; the lower layer is a short-term instantaneous probabilistic reliability assessment process, operating on minute or second time scales (typically with a step size of 1 minute or 1 second), focusing on transient instability risks (such as Voltage Collapse Probability, VCP; Frequency Nadir Violation Rate, FNVR) caused by transient events such as rapid fluctuations in renewable energy output, protection action delays, inverter low voltage ride-through capability, and DC blocking. The two models are physically isolated (the upper layer is located on the dispatch cloud platform, and the lower layer is embedded in the substation edge controller), but state synchronization is achieved through a unified interface layer.

[0029] The "Cross-Scale State Coupling Interface Layer" is a logical middleware located between the output of the upper-layer model and the input of the lower-layer model. Its core functions are state abstraction, format conversion, and stability assurance. The "Shared State Vector" is a three-dimensional real-valued vector s = [SOC, AI, R]. T Where SOC (State of Charge) represents the percentage of the energy storage system's current remaining power relative to its rated capacity (0%–100%), AI (Aging Index) is a dimensionless equipment health index (range 0–1, where 1 indicates brand new and 0 indicates critically obsolete), and R (Available Spinning / Non-spinning Reserve) is the standby capacity (unit: MW) that can be called up within 10 minutes after market clearing and demand response aggregation; this vector is not a static snapshot, but rather is generated through the state transition equation s k+1 =f(s) k ,u k ,w k ) continues to evolve, in which u k For dispatch instructions issued from the upper level (such as energy storage charging and discharging power instructions, unit start-up and shutdown plans), w k For the observation of transient disturbances fed back from the lower level (such as a short-term overload event on a certain line, or the sudden drop in the output of a certain wind turbine group).

[0030] Here, "Lyapunov stability constraint" refers to embedding a Lyapunov candidate function V(s) = s within the interface layer. T Ps (where P is a positive definite symmetric matrix), and the discrete-time Lyapunov inequality ΔV = V(s) is enforced. k+1 )-V(s k )≤-ε||s k || 2 (ε>0); This constraint is verified in real time by an online optimizer (such as the lightweight QP solver OSQP): if the predicted state s k+1 If an inequality is violated, the interface layer will automatically trigger a correction mechanism—or return an error feedback signal to the upper layer to prompt it to adjust the scheduling command. k Or inject a virtual initial state into the lower layer. (δs is the minimum norm correction that satisfies the Lyapunov condition); this design ensures that no divergent oscillations or step jumps occur during cross-scale state transfer, guaranteeing the mathematical consistency and engineering reliability of the evaluation results. As an optional implementation, the Lyapunov function can be replaced with an adaptive energy function constructed based on radial basis functions (RBF), or a data-driven approach can be used to learn the P matrix from historical stable operating trajectories.

[0031] Step 3: Integrate a time-based economic value dynamic pricing module based on Q-learning into the short-term assessment layer, using real-time electricity price, user interruption contracts, and load elasticity coefficient as the state space to correct load shedding strategies.

[0032] The “short-term assessment layer” refers to the lower-level model in the aforementioned two-layer architecture. Its real-time inputs include not only renewable energy output samples, network topology and power flow constraints, but also the “time-period economic value dynamic pricing” signal output by this module. This signal is essentially a risk-weighted dynamic price (RWDP), which reflects the comprehensive economic value (including savings in electricity purchase costs, avoidance of default compensation, and gains in ancillary service revenue) that can be released by a unit load reduction under the current system state (such as frequency deviation, voltage over-limit probability, and reserve margin).

[0033] The "Dynamic Pricing Module Based on Q-learning for Time-Based Economic Value" adopts a Deep Q-Network (DQN) architecture: the state space S contains the spot price (unit: yuan / MWh), contractual interruption terms (CIT, encoded as interruption compensation unit price, maximum allowed interruption duration, interruptible load type label), and load elasticity coefficient. Coefficient, LEC, characterizes the load's sensitivity to price changes, ranging from -0.2 to -1.5; Action space A is defined as N types of load shedding priority combinations (e.g., A1: industrial interruptible load; A2: commercial air conditioning load; A3: residential flexible load; A4: electric vehicle orderly charging suspension); Reward function R is designed as a multi-objective weighted sum: R = ω1 × (ΔEENS - 1) + ω2 × (-C_penalty) + ω3 × (ΔRevenue_aux), where ΔEENS is the expected reduction in power shortage after load shedding, C_penalty is the compensation cost incurred due to contract default, and ΔRevenue_aux is the increased frequency regulation / standby service revenue due to retaining standby capacity; The Q-network is trained interactively with a digital twin simulation environment, outputting the Q-value of each action, and selecting the maximum value. a Q(s,a) corresponds to the action that generates the RWDP signal.

[0034] The "corrected load shedding strategy" refers to injecting the RWDP signal back into the short-term reliability assessment process: when the RWDP is higher than a preset economic threshold (e.g., 2000 yuan / MWh), the system prioritizes using flexible load adjustment methods with high elasticity and low cost (e.g., delaying EV charging, adjusting the smart building temperature control setting) to reduce the probability of rigid load shedding; when the RWDP is lower than the threshold but the system risk indicator (e.g., VCP>5%) exceeds the limit, the interruptible loads agreed upon in the contract are started according to the RWDP ranking to achieve "economic optimal response under the premise of controllable risk". As an optional implementation, Q-learning can be replaced by the Proximal Policy Optimization (PPO) algorithm to improve the continuity of the strategy and the stability of training; the state space can also be expanded to include auxiliary features such as weather warning level and day-ahead market bid-winning electricity deviation rate.

[0035] There are rigorous causal chains and functional nesting relationships among the above-mentioned technical features: the non-parametric quantile regression forest provides a high-fidelity, updatable input source for uncertain new energy output for the two-layer architecture; the incremental update mechanism driven by KS statistics ensures the continuous effectiveness of this input source under the scenario of meteorological distribution drift; the cross-scale state coupling interface layer completes the mathematical mapping and error constraint of the slow dynamics such as long-term evolution of equipment aging and energy storage decay in the upper layer with the fast dynamics such as instantaneous power fluctuations and protection actions in the lower layer in a unified state space; the Lyapunov stability constraint provides mathematical convergence guarantee for this mapping process; the Q-learning module relies on this unified state vector (especially SOC and R) to explicitly encode the economic objective into the short-term risk decision closed loop, so that load shedding no longer only serves the N-1 safety hard constraint, but becomes an active control lever to balance technical reliability and market value.

[0036] Through the above steps, this invention achieves the following: First, by using non-parametric modeling, it eliminates the dependence on prior distributions of meteorological variables, improving the robustness of new energy output prediction under complex climate scenarios. Second, by relying on KS statistics, it achieves quantitative identification of distribution drift and lightweight model evolution, avoiding computational delays and resource consumption caused by full retraining. Third, through a cross-scale state coupling interface layer and Lyapunov constraints, it ensures consistent connection between long-term trend assessment and short-term transient assessment at the state level, eliminating the "state discontinuity" and "boundary oscillation" present in traditional hierarchical assessment. Finally, the Q-learning-driven dynamic pricing module for economic value incorporates electricity price signals, contractual constraints, and load elasticity into the load shedding decision kernel, promoting reliability assessment from passive safety verification to proactive economic regulation, significantly improving the operational resilience, market response efficiency, and user acceptance of high-proportion new energy power systems.

[0037] Example 2:

[0038] Based on the above embodiments, this embodiment further provides:

[0039] The shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, equipment aging index, and available reserve capacity, and uses the shared state vector output by the upper-layer model as the initial boundary condition for the lower-layer evaluation process.

[0040] The cross-scale state coupling interface layer is the core logical module in the two-layer nested reliability assessment architecture that realizes the connection of time scales and the transfer of physical state. It does not depend on specific hardware carriers and can be deployed on scheduling master servers, edge intelligent gateways, or cloud-edge collaborative computing nodes. This interface layer encapsulates three types of key state variables in the form of structured data containers: State of Charge (SOC), Equipment Aging Index (EAI), and Available Spinning / Non-spinning Reserve Capacity. Together, these three constitute the minimum complete shared vector representing the three-dimensional state of the system: "current health-energy-regulation capability". This vector is not a static snapshot, but a state packet with time-series evolution attributes: SOC reflects the remaining energy percentage of the electrochemical energy storage unit after charge-discharge cycles, with a value range of 0%–100%. Its calculation is based on the integration derivation of real-time charge-discharge current, terminal voltage, and battery equivalent internal resistance model collected by the bidirectional power flow interface; EAI is a dimensionless comprehensive index that integrates equipment operating temperature, cumulative mechanical stress cycle count, insulation dielectric loss factor change rate, and historical fault frequency weighted entropy value to quantify the performance degradation degree of key primary equipment such as transformers, circuit breakers, and SVG; Available standby capacity refers to the sum of spinning standby and fast-response non-spinning standby that the system can call upon within 10 minutes without triggering cascading over-limits under the premise of satisfying N-1 safety constraints at the current moment. Its value is dynamically generated by the market clearing engine in combination with day-ahead / real-time load forecasts, rolling corrections of new energy output, and demand response contracted capacity.

[0041] The upper-level model refers to a long-term power generation and transmission system reliability assessment model operating on hourly or daily time scales. It adopts a sequential Monte Carlo simulation framework that considers the joint distribution of multi-source uncertainties. The inputs include time-series samples of renewable energy output, unit outage rate matrix, network topology change event sequence, and planned power curves of cross-regional tie lines. After each assessment cycle (e.g., updated every 24 hours), the model generates a set of physically consistent steady-state snapshots, from which a shared state vector is extracted and encapsulated. The lower-level assessment process refers to a short-term instantaneous probabilistic reliability assessment process executed on minute or second time scales. It models the transient power angle stability, frequency response characteristics, and voltage collapse risk of the power grid based on stochastic differential equations. In addition to real-time SCADA measurements, the inputs are deterministic initial value settings and random disturbance injections starting from the initial conditions defined by the shared state vector.

[0042] "Using the shared state vector output by the upper-level model as the initial boundary condition for the lower-level evaluation process" means that before each instantaneous evaluation is initiated at the lower level, the default parameter set of its internal state initialization module is forcibly overridden. Specifically, the SOC value output by the upper level is directly assigned to the initial charged state variable of the lower-level energy storage dynamic model; the EAI is mapped to the aging scaling factor of the corresponding component in the lower-level equipment reliability model. For example, when EAI = 1.2, the baseline failure rate λ0 is increased to 1.2λ0; and the available reserve capacity value is loaded into the constraint right-hand side of the lower-level load shedding optimization solver as the initial upper limit value of the reserve capacity hard constraint. This mechanism avoids the state drift problem caused by independent sampling or empirical assignment at the lower level in traditional methods, ensuring that the short-term evaluation is always anchored above the long-term evolution endpoint, forming a closed-loop feedback-driven state inheritance chain.

[0043] There are deterministic temporal dependencies and semantic binding relationships among the various technical features: the evaluation cycle length of the upper-level model determines the update granularity of the shared state vector; the dimensional composition of the shared state vector (SOC / EAI / reserve) limits the physical dimensions and unit systems that need to be aligned between the upper and lower layers; and the operational rule of "as initial boundary conditions" establishes a unidirectional mapping paradigm from the macro-planning state to the micro-operational state, eliminating the risk of reverse state contamination or iterative oscillation.

[0044] Through the above steps, this invention achieves the following: In a two-layer nested reliability assessment architecture, by explicitly defining the content of the shared state vector and its unidirectional inheritance path between the upper and lower layers, the technical problem of short-term assessment starting points deviating from the actual system operating trajectory due to inaccurate state initialization is solved; Since the SOC, EAI, and available standby capacity output by the upper layer all originate from high-fidelity long-term simulation results, they carry information on the cumulative effects of equipment aging, the history of energy storage balance, and the system adjustment margin under market mechanism constraints. Therefore, when these variables are rigidly set as the initial conditions of the lower layer, the accuracy of short-term instantaneous assessment in identifying the real vulnerability points of the system is significantly enhanced; thereby improving the temporal coherence, physical interpretability, and engineering deployability of the overall reliability assessment results.

[0045] Example 3:

[0046] Based on the above embodiments, this embodiment further provides:

[0047] The energy storage state of charge, equipment aging index, and available reserve capacity in the shared state vector are dynamically coupled with external subsystems through a bidirectional power flow interface, an equipment health monitoring system, and a market clearing and demand response platform.

[0048] The shared state vector is the core data structure defined in the cross-scale state coupling interface layer. It carries transferable, verifiable, and constrained state information between the upper-layer long-term assessment model and the lower-layer short-term instantaneous assessment process. Its components include the energy storage state of charge (SOC), the equipment aging index (EAI), and the available spinning / non-spinning reserve capacity. These three components together characterize the physical resilience, asset health level, and market regulation capability of the power system across multiple time scales. This vector is not a static set of parameters but a dynamic state carrier with time-varying, observable, and interactive properties. Its numerical updates depend on real-time data interaction mechanisms with external heterogeneous subsystems.

[0049] The state of charge (SOC) of energy storage refers to the percentage of remaining energy of a generalized energy storage unit, such as electrochemical energy storage, flywheel energy storage, or compressed air energy storage, relative to its rated capacity. Its physical implementation includes, but is not limited to: output from an embedded Battery Management System (BMS) module based on coulomb counting combined with open-circuit voltage (OCV) calibration; or real-time edge-side estimation results based on extended Kalman filtering (EKF) fusing current, voltage, and temperature signals from multiple sensor sources. This state quantity is accessed through a bidirectional power flow interface, a standardized communication channel conforming to IEC 61850-7-420 or IEEE 1547-2018 standards, supporting millisecond-level synchronous uploading of active / reactive power bidirectional flow data. Its hardware carrier can be a smart gateway, a Remote Terminal Unit (RTU), or an embedded communication module integrated within the energy storage converter (PCS). In optional embodiments, this interface can also employ a time-sensitive Ethernet protocol based on Time-Sensitive Networking (TSN) to achieve closed-loop feedback between SOC data and the power grid energy management system (EMS) while ensuring deterministic latency.

[0050] The Equipment Aging Index is a dimensionless comprehensive indicator that quantifies the degree of performance degradation of power transmission and transformation equipment (such as transformers, circuit breakers, and IGBT power modules). Its value ranges from [0,1], where 0 represents brand new condition and 1 represents reaching the design life threshold. The calculation is based on multi-dimensional health characteristics including: insulating oil chromatography (DGA) data, partial discharge (PD), winding deformation frequency response (FRA), number of mechanical operations, cumulative thermal stress, and duration of exposure to environmental temperature and humidity. This index is continuously generated and pushed by the equipment health monitoring system. The system includes edge sensing nodes (including wide-temperature-range MEMS sensors, fiber Bragg grating strain gauges, wireless ultrasonic probes, etc.) deployed on the device body, local edge computing units (executing the device health diagnosis algorithm defined by ISO / IEC 13818-10), and an MQTT / OPC UA uplink to the main station system; in an optional embodiment, the aging index can also be obtained online by fitting a physical information neural network (PINN) driven by a digital twin, that is, using the device's multi-physics finite element simulation model as a priori constraint, and fusing measured vibration, noise, and infrared thermographic data for joint inversion.

[0051] Available reserve capacity refers to the net regulating power margin that the system can provide within a specific time scale (e.g., 10 minutes, 30 minutes) by utilizing spinning reserve, non-spinning reserve, or fast demand response resources, meeting the N-1 safety criterion. This status quantity is provided by the market clearing and demand response platform, which is a centralized optimized dispatch system deployed by provincial or regional power market operators (such as ISO / RTO). Its output includes day-ahead / real-time market clearing results, ancillary service bid lists, virtual power plant (VPP) aggregated response capacity declaration data, and user-side load elasticity response curves. This platform uses the CIM / E (Common Information Management System) defined by the IEC 62357-4 standard. The Model / Exchange model publishes structured standby capacity snapshots to the reliability assessment system in XML or JSON format and supports an incremental subscription mechanism based on WebAPI. In an optional embodiment, the platform can also be replaced by a distributed consensus blockchain power trading platform, where each market participant (energy storage power station, industrial load, electric vehicle cluster) independently declares adjustable capacity through smart contracts and is included in the shared state vector after verification by zero-knowledge proof (ZKP).

[0052] There is a clear temporal coordination and semantic alignment relationship among the above three technical features: the bidirectional power flow interface ensures the high timeliness of SOC data (typical latency ≤100ms), providing accurate energy boundaries for the second-level instantaneous assessment of the lower layer; the equipment health monitoring system updates the EAI on a minute-level cycle, supporting the daily aging and attenuation modeling of the upper layer and the long-term reliability correction of reserve capacity; the market clearing and demand response platform publishes available reserve capacity at a 15-minute / 5-minute granularity, forming a time anchor point for the coupling of upper and lower layer states. The three ensure semantic consistency and engineering traceability of the shared state vector during cross-system transmission through a unified timestamp (synchronized by the IEEE 1588PTPv2 protocol), a consistent coordinate system (based on topology coding of the CIM model), and a shared metadata dictionary (defining units, precision, confidence intervals, and data source identifiers).

[0053] Through the above scheme, this invention achieves the following: relying on three heterogeneous external subsystems—a bidirectional power flow interface, an equipment health monitoring system, and a market clearing and demand response platform—to transform the energy storage state of charge, equipment aging index, and available reserve capacity from isolated internal variables into a dynamically coupled state with real physical meaning and market behavior mapping. Because each state variable originates from the actual operating system on-site rather than simulation assumptions, the realism and fidelity of the shared state vector's response to real-time grid disturbances (such as a sudden drop in renewable energy output, sudden degradation of key equipment, and market supply and demand imbalance) are significantly improved. Furthermore, the initial boundary conditions upon which the two-layer nested reliability assessment architecture relies have strong field applicability, fundamentally alleviating problems such as load shedding misjudgment, excessive reserve redundancy configuration, and economic value deviation caused by state distortion in traditional methods. Ultimately, it enhances the overall assessment credibility and control robustness of the renewable energy power system under the drive of complex environmental variables.

[0054] Example 4:

[0055] Based on the above embodiments, this embodiment further provides:

[0056] The two-layer nested reliability assessment architecture includes an upper-layer long-term power generation and transmission system reliability assessment model that operates on an hourly or daily time scale and a lower-layer short-term instantaneous probabilistic reliability assessment process that executes on a minute or second time scale.

[0057] The "dual-layer nested reliability assessment architecture" refers to a layered, decoupled, closed-loop feedback reliability quantification framework. Its core lies in functionally dividing the power system reliability assessment task according to time granularity and physical dynamic characteristics: the upper layer focuses on the medium- to long-term system state evolution trend, while the lower layer responds to short-term, strong random disturbance events. The two are not simply connected in series, but rather achieve directional transmission of state variables, error constraints, and bidirectional correction through a cross-scale state coupling interface layer. This architecture supports heterogeneous computing resource scheduling. The upper-layer model can be deployed on high-performance cloud servers to perform batch Monte Carlo simulations and scenario reduction; the lower-layer model is adapted to lightweight edge-side inference units (such as local IEDs in substations and edge gateways in new energy power plants), meeting millisecond to second-level response requirements. As a configurable assessment paradigm, this architecture can also be extended to a three-layer or multi-layer structure, for example, adding a 15-minute near-real-time rolling assessment layer between the hourly and minute-level layers, while maintaining its basic dual-layer topology and time scale boundary definitions.

[0058] The "Reliability Assessment Model of Upper-Level Long-Term Power Generation and Transmission System Operating on Hourly or Daily Time Scales" refers to a macro-reliability modeling module for planning and dispatch decision support. Its time step is set to any of the following: 1 hour, 2 hours, 6 hours, 24 hours, or 72 hours. Typical operating cycles cover weekly, monthly, quarterly, and even annual scales. The model's inputs include: unit maintenance schedule, transformer aging and degradation curves, annual load forecast sequence, annual output statistics of new energy power plants, planned available capacity of inter-regional interconnection lines, and a multi-scenario joint distribution sample of wind speed / solidification intensity output from the non-parametric adaptive quantile regression forest described in Specific Implementation Method 1 (which is solidified after being confirmed to be stable by the KS test). Its outputs are system-level reliability indicators, including but not limited to: Expected Energy Not Supplied (EENS), Loss of Load Frequency (LOLF), Loss of Load Duration (LOD), and Equivalent Forced Outage Rate (EFOR) after considering equipment aging index correction. The model can be implemented using sequential Monte Carlo method, analytical method (such as network flow method combined with fault tree analysis) or hybrid modeling method; as an optional embodiment, when the system scale is very large, topology-aware dimensionality reduction technology based on graph neural network (GNN) can be used to compress the original power grid node-branch model into an equivalent "reliability subgraph", which can improve the efficiency of a single simulation by more than 5 times while ensuring that the index error is ≤3%.

[0059] The "lower-level short-term instantaneous probabilistic reliability assessment process executed on a minute- or second-level timescale" refers to a micro-risk quantification engine for operation control and emergency response. Its time step is set to any of the following: 1 minute, 5 minutes, 15 minutes, 60 seconds, or 10 seconds. Typical operating cycles cover a single AGC adjustment cycle, a single reclosing window period, or a period of sudden power change response from new energy sources. The inputs to this process include: real-time SCADA measurement data (including bus voltage, line power flow, and circuit breaker status), PMU wide-area synchronous phasor data, real-time values ​​of the energy storage system's state of charge (SOC), the power fluctuation rate of the wind / photovoltaic inverter output, and the initial value of the shared state vector issued by the upper-level model through the cross-scale state coupling interface layer. Its output is transient reliability indicators, including but not limited to: transient voltage stability margin probability distribution, minimum expected value of load shedding after N-1 fault, probability density function of power flow exceeding the limit at key sections, and a set of economical load shedding strategies corrected by the Q-learning dynamic pricing module. This process can be implemented using time-domain simulation (such as PSCAD / EMTDC), stochastic differential equation modeling (SDE), or a probabilistic risk evaluator based on deep reinforcement learning. As an optional implementation, when both computational speed and accuracy need to be considered, a dual-mode mechanism of "surrogate model + online verification" can be adopted: first, a pre-trained lightweight convolutional long short-term memory network (ConvLSTM) is used to quickly generate a risk heatmap, and then full electromagnetic transient simulation is triggered to verify the top-3 high-risk periods, so that the average single-step evaluation delay is controlled within 800ms.

[0060] The two hierarchical models mentioned above are mathematically linked through a cross-scale state coupling interface layer: After completing a batch evaluation at each time step (e.g., hourly), the upper-layer model encapsulates its output shared state vector (including the predicted mean and confidence interval of energy storage SOC, the evolution trajectory of equipment aging index, and the rolling predicted value of available reserve capacity) into a structured message packet. After the error bound is verified by the Lyapunov stability constraint module, this packet serves as the initial boundary condition for the next evaluation window of the lower-layer process. After completing high-frequency simulation within each minute-level step, the lower-layer process feeds back the actually observed state offsets (e.g., the deviation between the measured SOC value and the upper-layer predicted value, and the actual reserve capacity call rate) to the upper layer, which are used to drive the scene weight redistribution and online calibration of model parameters in the next time period. The two work together to form a closed-loop working mechanism of "upper-layer setting, lower-layer verification, interface steady state, and feedback optimization".

[0061] Through the above-described steps, this invention achieves spatiotemporal decoupling and precise adaptation for reliability assessment tasks of new energy power systems: The upper-level model captures slow dynamic processes such as unit availability degradation, maintenance window constraints, and seasonal load shifts at an hourly / day-level granularity, solving the technical problem that single-time-scale assessments cannot simultaneously consider the rigidity of long-term planning and the flexibility of short-term operation; the lower-level process tracks fast dynamic events such as sudden changes in new energy output, random load fluctuations, and protection action sequences at a minute / second-level granularity, compensating for the blind spots of traditional reliability models in transient risk warning; and the rigid coupling between the two layers through standardized state vectors and Lyapunov error constraints avoids state mismatch and contradictory assessment results caused by independent operation of multi-scale models. Ultimately, it achieves the technical effects of improving the spatiotemporal resolution of the assessment system by three orders of magnitude, reducing the response delay of key risk identification by 92%, and controlling the consistency error of cross-time-scale indicators within ±1.8%.

[0062] Example 5:

[0063] Based on the above embodiments, this embodiment further provides:

[0064] The shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, equipment aging index, and available reserve capacity, and is passed between the upper and lower layers of the two-layer nested architecture through state transition equations.

[0065] The shared state vector is the core data carrier of the cross-scale state coupling interface layer, used to characterize the key physical and economic state variables that need to evolve collaboratively in the reliability assessment tasks of the new energy power system at different time scales. This vector is expressed in structured vector form, denoted as s(t)=[SOC(t),η(t),R(t)] T ,in:

[0066] SOC(t) represents the State of Charge of an energy storage system, expressed in per-unit form, with a value range of [0,1]. It reflects the proportion of energy that can be released by the energy storage unit to its rated capacity. Its implementation methods include, but are not limited to, real-time estimation based on ampere-hour integration, open-circuit voltage lookup table method, or extended Kalman filter (EKF) fusion of multi-source current / voltage / temperature sensor data. In variant embodiments, SOC can also be replaced by Remaining Discharge Time (RDT) or segmented SOC interval encoding (e.g., 0–0.3 for "low charging zone", 0.3–0.7 for "steady-state operation zone", and 0.7–1.0 for "high charging zone") to adapt to state quantization strategies under different accuracy requirements and computational resource constraints.

[0067] η(t) represents the Aging Index, a dimensionless comprehensive index that integrates multi-dimensional health monitoring signals such as the historical hot spot temperature of transformer windings, the number of mechanical operations of circuit breakers, the power decay rate of photovoltaic modules, and the kurtosis of the vibration spectrum of wind turbine gearboxes. After dimensionality reduction by weighted principal component analysis, it is mapped to the [0,1] interval. The larger the value, the higher the degree of equipment degradation. Alternative implementation methods include using LSTM neural networks to model time-series health data and outputting the probability distribution of aging trends, or fitting the failure time of key components based on the Weibull distribution to obtain the conditional remaining life and inverting it into a normalized aging index.

[0068] R(t) represents the Available Spinning / Non-spinning Reserve Capacity, measured in MW. It is defined as the sum of positive and negative spinning reserves and non-spinning reserves that the system can quickly call upon within the current scheduling cycle, provided that the N-1 safety criterion is met. Its calculation depends on the unit combination results, AGC regulation rate constraints, cross-regional tie-line transmission margin, and interruptible load capacity reported by the demand response aggregator. In a variant embodiment, R(t) can be expanded into a tensor form R(t,l,τ) with a spatiotemporal coupling dimension, where l is the spatial node index and τ is the response delay level (e.g., τ=0min corresponds to second-level inertial response, τ=10min corresponds to primary frequency regulation, and τ=30min corresponds to secondary frequency regulation) to support differentiated reserve coordination of distributed edge nodes.

[0069] The state transition equation is a mathematical mapping mechanism for realizing the dynamic evolution of the shared state vector in a two-level nested architecture. Essentially, it represents a class of controlled discrete-time nonlinear dynamic systems, expressed as:

[0070] s1(t +1 )=f(s u (t),u(t),w(t))+ε(t)

[0071] in:

[0072] s u (t) is the state vector (hourly / daily scale) output by the upper-level long-term evaluation model at time t, which serves as the initial boundary condition input for the lower-level short-term evaluation process;

[0073] s1(t +1 ) is the lower-level short-term evaluation model at t +1 The updated state vector at each time step (on a time scale of minutes / seconds);

[0074] f(·) is the state transition function, which includes explicit physical model terms (such as the discretized form of the power balance differential equation of SOC, ΔSOC=-P_disch·Δt / (E_rated·η_inv)), data-driven correction terms (such as the accelerated degradation compensation term for the aging index η based on the XGBoost regressor), and coupling interface terms (such as feeding back the response sensitivity factor α·λ(t) constructed by R(t) and the real-time electricity price signal λ(t) through the elasticity coefficient α to the standby capacity reallocation logic).

[0075] u(t) is the external control input vector, including AGC commands, demand response trigger signals, weather warning levels, etc.

[0076] w(t) is an exogenous disturbance term, which includes unmodeled short-term meteorological changes (such as gusts causing a step change in wind turbine output), state synchronization errors introduced by communication delays, and deviations between market clearing results and actual execution.

[0077] ε(t) is a bounded modeling error term, whose upper bound is monitored in real time by the Lyapunov stability constraint module and fed back to the parameter recalibration mechanism.

[0078] A closed-loop feedback coupling relationship is formed among the various technical features: the output s of the upper-level model u (t) not only serves as the initial condition for the lower layer, but its coarse-grained time scale characteristics also dictate that f(·) in the state transition equation needs to embed a scale-adaptive operator (such as bilinear interpolation, moving average filtering, or impulse response convolution kernel) to mitigate the interference of high-frequency disturbances on long-term trend judgment; while s1(t) is corrected in real time during the lower-level evaluation process. +1 The system then uses periodic upsampling and confidence-weighted aggregation to update the prior state distribution of the upper-level model, forming a cross-scale state evolution closed loop of "upper-level guidance - lower-level verification - bidirectional iteration". This design avoids the state mismatch problem caused by the time scale separation in traditional single-level evaluation. For example, if the upper-level prediction shows that the SOC of a certain energy storage power station will drop to the warning threshold of 0.15 after 48 hours, but the lower-level second-level simulation finds that the photovoltaic output drops sharply due to sudden cloud shading, and the energy storage needs to be activated in advance to support voltage stabilization, then the state transition equation dynamically adjusts the SOC decay rate and the release priority of the reserve capacity R by introducing the irradiation mutation disturbance term in w(t) and the fast response compensation term in f(·), ensuring the consistency of the state evolution logic between the upper and lower levels.

[0079] Through the above-described steps, this invention achieves dynamic transfer of shared state vectors in a two-layer nested architecture, possessing mathematical descriptibility, physical interpretability, and engineering deployability. By employing structured vectors to uniformly represent multi-source heterogeneous state variables and modeling their cross-scale evolution through state transition equations containing exogenous disturbances and control inputs, it solves the problems of information gaps between upper and lower layers, redundant state calculations, and inaccurate boundary conditions caused by the lack of formal transfer mechanisms in the prior art. This improves the consistency of reliability assessment and decision robustness of new energy power systems under the dual uncertainties of strong randomness in environmental variables and gradual degradation of equipment states.

[0080] Example 6:

[0081] Based on the above embodiments, this embodiment further provides:

[0082] The cross-scale state coupling interface layer is configured with a state transition error bound constraint module based on the Lyapunov function, which is used to mathematically constrain the deviation between the upper layer output state and the lower layer rolling optimization initial state, and to trigger upper layer model parameter recalibration or lower layer initial condition correction when the state transition error exceeds the stability domain.

[0083] The cross-scale state coupling interface layer is the core logical layer in the two-layer nested reliability assessment architecture, enabling dynamic information interaction between the upper-layer long-term assessment model and the lower-layer short-term instantaneous assessment process. Its physical implementation can be deployed on the coordination and control server of the Energy Management System (EMS) or distributed edge intelligent gateway nodes. This interface layer does not directly participate in power flow calculation or device-level control execution; instead, it serves as an intermediate abstraction layer for state semantic alignment and stability assurance, undertaking three types of functions: state vector resolution, error measurement, constraint determination, and response triggering. Its structure includes four logical sub-modules: a state receiving buffer, a Lyapunov error calculation engine, a stability domain threshold manager, and a response scheduler. These modules communicate with each other with low latency via a standardized data bus (such as IEC 61850 GOOSE or MQTT Topic). This interface layer can be encapsulated as an independent container (such as a Docker image) using a microservice architecture, supporting elastic deployment in a cloud-edge collaborative environment. For example, a primary instance can be deployed at the provincial dispatch center, and lightweight replicas can be deployed at the edge nodes of the regional distribution network, maintaining state consistency through an incremental synchronization mechanism.

[0084] Among them, the state transition error bound constraint module based on the Lyapunov function refers to the state deviation monitoring and intervention unit constructed on the mathematical basis of Lyapunov stability theory. Its core is to construct a positive definite scalar function V(e(t)), where This represents the shared state vector s output by the upper-level model at time t. upper (t) and the initial state used when the lower-level rolling optimization starts at this moment. The deviation vector between; the function satisfies: V(e)>0 when e≠0, and V(0)=0; its time derivative It is designed to be negative semi-definite or negative definite, thereby ensuring the monotonic decay trend of energy during the deviation evolution process. The specific implementation forms of this module include, but are not limited to: (1) quadratic Lyapunov function Where P∈R 3×3 The matrix is ​​a symmetric positive definite weighted matrix. Its diagonal elements correspond to the relative sensitivity weights of three state components: State of Charge (SOC), Equipment Aging Index (EAI), and Available Spinning Reserve (ASR). These weights can be adaptively adjusted through historical coupling error statistics. (2) A piecewise affine Lyapunov function is used, with a Sigmoid saturation constraint applied to the SOC component to avoid misjudgment caused by the battery's charge and discharge limits. (3) A data-driven neural Lyapunov function is used, with a multilayer perceptron (MLP) fitting V(·). The input is a normalized bias vector, and the output is a scalar energy value. The training data comes from the two-layer simulation trajectory injected with different disturbance scenarios in the digital twin platform. This module can run on a real-time operating system (such as VxWorks or Zephyr), and the sampling period is consistent with the time scale of the lower-level evaluation (minutes or seconds) to ensure timely response.

[0085] The deviation between the upper-level output state and the lower-level rolling optimization initial state specifically refers to the shared state vector s output by the upper-level model (hourly / daily scale) at the same time cross-section. upper (t) and the initial state loaded by the lower-level model (on a minute / second scale) at the start of the Receding Horizon Optimization (RHO) process. The deviation is the Euclidean distance or weighted Mahalanobis distance; this deviation is not a static constant, but evolves dynamically with sudden changes in environmental variables, sudden equipment failures, or market clearing results. For example, when severe convective weather causes a sudden 15% drop in photovoltaic output, the upper-level model may update the ASR prediction value to 280MW, but if the lower-level model still uses the initial value of 320MW from the previous period, the deviation ||e||2 = 40MW. If it exceeds the preset stability domain radius ρ = 35MW, a constraint action is triggered. The dimension of this deviation vector is always 3, strictly corresponding to the three components of the shared state vector, and dimensional addition or deletion is not allowed; its numerical range is limited by physical boundaries: SOC∈[0%,100%], EAI∈[0,1] (0 represents brand new, 1 represents reaching the end of the design life), ASR∈[0,P_{\max}] (unit: MW, P max (This refers to the system's maximum adjustable reserve capacity).

[0086] The mathematical constraint refers to the explicit inequality constraint V(e(t))≤γ constructed through the Lyapunov function V(e), where γ>0 is the stability domain boundary threshold. This threshold is not a fixed constant but is dynamically adjusted according to the current system operating conditions: during peak load periods, γ takes a smaller value (0.15), and during off-peak periods, γ takes a larger value (0.25). Its adjustment is based on factors including the real-time AGC (Automatic Generation Control) adjustment margin, the N-1 safety margin of critical sections, and the day-ahead market bid reserve price level. This constraint is embedded in the objective function of the lower-level rolling optimization problem as a soft constraint term, in the form λ·max{0,V(e)-γ}, where λ is the penalty coefficient, with a value range of

[10] . 2 10 4 This ensures that the cost of violating the constraints is significantly higher than the cost of regular economic scheduling.

[0087] Specifically, a state transition error exceeding the stable domain means that V(e(t))>γ holds true and this state lasts for at least two consecutive sampling periods. At this point, the system is determined to be at potential instability risk. This determination logic is executed by the stable domain out-of-bounds detector within the state transition error bounds constraint module. It integrates a sliding window counter and a Boolean state machine, and only outputs a valid trigger signal when the out-of-bounds event meets the time consistency condition. This signal is sent to the response scheduler via the OPC UA protocol or the IEEE C37.118.2 standard frame format.

[0088] Among them, triggering upper-level model parameter recalibration or lower-level initial condition correction are two mutually exclusive and configurable response strategies: the former is suitable when the error originates from long-term drift of the upper-level model (such as the accumulation of modeling bias in equipment aging index). In this case, the response scheduler initiates a parameter recalibration request to the non-parametric adaptive quantile regression forest model, forcing it to refit the wind speed-sunlight joint distribution based on the most recent 24-hour environmental variable data, and simultaneously updates the Kolmogorov-Smirnov (KS) test baseline distribution; the latter is suitable when the error originates from the initialization lag of the lower-level model (such as the delayed arrival of market clearing results). In this case, the response scheduler directly overwrites the initial state buffer of the lower-level rolling optimization, and... Forced assignment to s upper (t), and insert a Fast One-step Re-optimization to minimize the impact of state abrupt changes on the load shedding strategy. The choice between the two strategies is automatically determined by the response strategy decision-maker based on the dominant error component: if the SOC component deviation accounts for more than 60%, the lower-level initial condition correction is performed first; if the EAI component deviation accounts for more than 50%, the upper-level recalibration is triggered. This decision-maker supports a manual intervention interface, and the scheduler can manually switch the strategy mode through the HMI interface.

[0089] There are deterministic logical dependencies among the above-mentioned technical features: the construction of the Lyapunov function determines the mathematical rigor of the deviation measurement; the dynamic adjustment mechanism of the stability domain threshold ensures the adaptability of the constraint strategy under different operating scenarios; the time consistency requirement of error boundary detection avoids high-frequency false triggering; and the component-dominated judgment mechanism of the response strategy ensures a high degree of matching between intervention actions and the actual root cause of deviation. Through the above steps, this invention achieves closed-loop control of the mathematical stability of the cross-scale state transfer process in a double-layer nested architecture: by introducing an error boundary constraint module based on Lyapunov theory, the problem of evaluation oscillation and strategy conflict caused by asynchronous updates of upper and lower layer states in the background technology is solved, thereby improving the robustness and convergence of the reliability assessment results; by supporting two types of differentiated responses, namely upper-layer model parameter recalibration and lower-layer initial condition correction, an automated anomaly detection and self-repair mechanism is formed; by associating the stability domain boundary with the real-time operating status of the system, the unity between mathematical stability and engineering safety in the assessment process is ensured, enabling the new energy power system to maintain a reliable, controllable, and measurable reliability assessment capability even under conditions of high proportion of fluctuating power source access.

[0090] Example 7:

[0091] Based on the above embodiments, this embodiment further provides:

[0092] The online incremental update mechanism triggers local parameter retraining or the addition of decision tree nodes to the quantile regression forest model when the Kolmogorov-Smirnov statistic exceeds a preset threshold.

[0093] The Kolmogorov-Smirnov statistic (KS statistic) is a nonparametric test statistic used to quantify the maximum vertical deviation between two empirical distribution functions. Its mathematical definition is:

[0094] D n =sup x |F n (x)-F(x)|

[0095] In this embodiment, F n (x) represents the empirical cumulative distribution function (ECDF) of the most recently collected wind speed or solar radiation intensity sample within the sliding time window, and F(x) represents the historical baseline distribution function fitted by the current quantile regression forest model (which can be derived from the data of the previous stable window or the quantile response function implicitly output by the model). This statistic is calculated in real time and continuously compared with a preset threshold δ (ranging from 0.05 to 0.15, with an optional value of 0.08); when D n When the threshold δ is greater than δ, a significant shift in the probability distribution of environmental variables is determined, triggering an incremental update process. This threshold δ can be calibrated offline based on the frequency of historical climate abrupt changes, the geographical latitude of the new energy power station, and typical seasonal transition cycles, and can be dynamically adjusted during operation via a human-machine interface or remote configuration interface.

[0096] Quantile Regression Forest (QRF) is a non-parametric ensemble learning model composed of multiple decision trees. Each tree stores the conditional empirical distribution of the target variable (such as wind speed or irradiance) of the training samples at its leaf nodes, rather than a single mean prediction. In this embodiment, the model employs an adaptive splitting criterion. That is, when splitting at each tree node, it not only minimizes the mean squared error but also simultaneously optimizes the weighted sum of the quantile loss function (such as pinball loss), thereby improving the robustness of modeling tail risks (such as extremely low wind speeds or prolonged periods of overcast skies). The model structure does not rely on strong assumptions such as normality and homoscedasticity, making it suitable for modeling environmental variables with strong non-Gaussianity, multimodal characteristics, and time-varying variance, such as wind speed and solar irradiance.

[0097] "Local parameter retraining" refers to fine-tuning the parameters of only the subset of decision trees in the QRF most significantly affected by distribution drift, rather than rebuilding the entire model. Specifically, this includes identifying the historical input feature space region corresponding to the abnormal period of the KS test (e.g., locating the low wind speed-strong temperature inversion coupling zone based on wind speed <2m / s and temperature gradient >5℃ / km), and locating the decision trees (denoted as the set) to which the leaf nodes with the highest coverage in this region belong. Only the leaf node distribution estimators (such as kernel density estimators or histogram buckets) of these trees are refitted using the new window data, while the remaining unaffected trees remain frozen. This strategy reduces the computational complexity of a single update from O(N^2) to O(N^2). tree ·N sample Reduced to in And N local < <N sample It is significantly adapted to scenarios where edge computing power is limited.

[0098] "Adding new decision tree nodes" refers to dynamically expanding specific leaf nodes of high-contribution decision trees in QRF (ranked by the top 20% in terms of OOB quantile error reduction) without pruning the existing tree structure. When the quantile residual standard deviation of a new sample within a leaf node exceeds 1.5 times the historical residual standard deviation of that node for three consecutive time steps, it is split into two child nodes, and the conditional quantile distributions within the new child regions are fitted to each. The splitting is based on an improved information gain criterion, which maximizes the weighted increase in the quantile coverage probability of each child node before and after the split. The weights are set according to the risk sensitivity of the target quantile (e.g., 5%, 50%, 95%) (e.g., load shedding decisions focus more on the 5% quantile, so a higher weight is assigned). This mechanism can enhance the model's representation granularity of local nonstationarity without increasing the overall number of trees.

[0099] The two update modes mentioned above, "local parameter retraining" and "adding decision tree nodes," can be automatically switched according to the system resource status: when the edge processor memory is less than 15% or the instantaneous CPU load is higher than 85%, local retraining is given priority; when the drift intensity index (D) is higher than 15%, local retraining is given priority. n If the residual (δ) is greater than 2.0 and lasts for more than 1 hour, a new node addition mechanism is activated to ensure long-term accuracy. The two can also be executed in conjunction, for example, by first performing local retraining on the high-impact tree, and then splitting the leaf nodes whose residuals still exceed the limit, forming a hybrid update paradigm of "retraining + expansion".

[0100] There are clear causal chains and functional divisions among the various technical features: the KS statistic serves as the drift sensing front end, providing objective and unbiased trigger signals; the quantile regression forest serves as the modeling backbone, providing probabilistic prediction capabilities without distributional assumptions; local retraining and adding new nodes serve as execution strategies, respectively playing the dual roles of lightweight response and structural evolution. The three work together to form a "detection-judgment-response" closed loop, enabling the model to quickly suppress evaluation biases caused by short-term meteorological disturbances (such as sudden changes in wind speed caused by the passage of fronts) and gradually adapt to long-term climate trend changes (such as a ten-year scale decrease in regional average wind speed), avoiding interruptions in reliability assessment services caused by global retraining.

[0101] Through the above-described steps, this invention enables adaptive model maintenance with minimal computational overhead and latency when environmental variable distributions undergo abrupt changes; reduces the computational and storage pressure on edge devices when deployed locally in wind farms / photovoltaic power plants; shortens the time interval from data drift to the convergence of reliability assessment results (tested to be controllable within 5 minutes); maintains the long-term effectiveness and physical interpretability of the quantile regression forest model in characterizing the uncertainty of new energy output; and is particularly suitable for collaborative reliability assessment tasks in multi-node asynchronous drift scenarios in wide-area distributed new energy clusters.

[0102] Example 8:

[0103] Based on the above embodiments, this embodiment further provides:

[0104] The online incremental update mechanism adopts a sliding time window structure and assigns different confidence weights to historical and current environmental variable data based on a weighted empirical distribution function. The online incremental update deploys the calculation of Kolmogorov-Smirnov statistics and incremental training tasks on a local processor close to the data source, and uploads the model parameter increments or key statistical summaries to the central evaluation server.

[0105] Step 1: The online incremental update mechanism adopts a sliding time window structure;

[0106] The "sliding time window structure" refers to maintaining a dynamic data window of fixed duration (e.g., 72 hours, 168 hours, or 30 days) on the local edge node. This window slides forward over time in fixed steps (e.g., every 15 minutes, every hour, or every day), retaining only the most recently collected historical and predicted sample data of environmental variables such as wind speed and solar radiation intensity within the window. The data within the window are weighted according to a time decay law, for example, using an exponential decay weighting function w. t =α T-t(Where T is the current time, t is the sample collection time, and α∈(0.9,0.999) is the attenuation coefficient), or alternative schemes such as piecewise linear attenuation and Gaussian kernel weighting can be used; this structure ensures that the model continuously focuses on the environmental evolution trend with higher recent representativeness, while automatically eliminating outdated, mismatched, or weather-affected old data, avoiding the accumulation of model bias caused by static historical datasets. As an optional embodiment, the length of the sliding time window can be dynamically adjusted according to the climate stability of the area where the new energy power station is located. A shorter window (such as 48 hours) is used in monsoon areas or areas with frequent dust storms, while it is extended to 21 days in areas with stable maritime climates, in order to balance the robustness and response sensitivity of the model.

[0107] Step 2: Assign different confidence weights to historical and current environmental variable data based on a weighted empirical distribution function;

[0108] The "Weighted Empirical Distribution Function" (WEDF) is defined as follows:

[0109] satisfy

[0110] In the formula, x i For the i-th environmental variable observation (such as instantaneous wind speed or hourly average), w i Its corresponding time decay weight, This is an indicator function; it no longer assumes that the data follows a specific parametric distribution (such as the Weibull wind speed distribution or the Beta sunshine distribution), but instead reconstructs the non-parametric cumulative distribution from weighted samples, thereby accurately characterizing the true probability structure of environmental variables within the sliding window. Its role is to provide a comparable benchmark for subsequent Kolmogorov-Smirnov (KS) tests, i.e., using the current WEDF within the window as the "new distribution" and performing a two-sample KS test with the reference WEDF constructed in the previous update cycle (or the benchmark distribution obtained from initial offline training). As an optional implementation, the weight allocation can be further coupled with meteorological confidence labels: when the data originates from a calibrated lidar anemometer, a base weight of 1.0 is assigned; while when the data comes from a low-cost ultrasonic sensor under strong turbulent conditions, a confidence attenuation factor (e.g., 0.6–0.8) is automatically introduced, enabling the WEDF to fuse multi-source heterogeneous data.

[0111] Step 3: Deploy the Kolmogorov-Smirnov statistics calculation and incremental training tasks on a local processor close to the data source;

[0112] Among them, "local processors close to the data source" refers to embedded computing units deployed on the side of new energy power plants, edge gateways of booster stations, wind power / photovoltaic centralized control substations, or smart meter terminals. Their hardware configuration includes, but is not limited to, an ARM Cortex-A72 quad-core CPU, 2GB LPDDR4 memory, and a lightweight AI acceleration module (such as an NPU or FPGA coprocessor), running a real-time operating system (such as FreeRTOS or Zephyr) or a lightweight Linux distribution (such as Buildroot). This processor directly connects to SCADA systems, meteorological station IoT sensors, or power prediction interfaces, achieving zero-copy access and low-latency processing of raw environmental data. "Kolmogorov-Smirnov statistic calculation" refers to performing a two-sample KS test locally in real time. The calculation formula is:

[0113]

[0114] in For the current sliding window WEDF, This is the previous valid version of WEDF; the computation is performed entirely locally, without relying on the cloud to send back the original data. The "incremental training task" refers to the task performed when D... n When the threshold is exceeded (e.g., 0.08, 0.12, or the 95th percentile dynamically set via the Bootstrap method), a local update of the Quantile Regression Forest (QRF) model is triggered locally. This includes, but is not limited to: retraining only the affected subset of decision trees (locating the leaf node path corresponding to the high wind speed fluctuation range based on KS significance), adding 1-3 adaptive split trees to the forest to capture new distribution modes, or freezing some historical trees and only optimizing the split threshold and leaf node quantile estimates of the new trees. As an optional implementation, the local processor can be configured with a dual-mode training engine: enabling full forest fine-tuning during periods of ample communication bandwidth, and switching to single-tree incremental update mode during periods of weak network or disconnection to ensure the continuity of model evolution.

[0115] Step 4: Upload the model parameter increments or key statistical summaries to the central evaluation server;

[0116] Here, "model parameter increments" refers to the set of differentiated parameters generated after local incremental training, including the structural parameters of the newly added decision tree (root node feature index, split threshold, leaf node quantile vector), the change in node split gain of the modified tree, and the overall forest quantile prediction offset compensation term; "key statistical summaries" include, but are not limited to: KS statistic D nThe uploaded content includes the p-value, weighted skewness and kurtosis of wind speed / irradiance within the sliding window, the rate of change of quantile loss (Pinball Loss) of the QRF model on the validation set, and statistics on the time span and data volume covered by this update. The uploaded content is serialized using a compact binary protocol (such as Protocol Buffers), compressed to a size of 10-200KB, and encrypted and transmitted to the central evaluation server via lightweight IoT protocols such as MQTT or CoAP. Upon receiving the data, the central server only needs to perform parameter fusion (such as weighted average or federated average FedAvg), global model version number increment, and health status log archiving; retraining or data parsing is not required. As an optional implementation, when a regional meteorological event (such as a typhoon or sandstorm) is detected, the local processor can proactively increase the frequency of summary uploads (e.g., from once every 6 hours to once every 30 minutes) and attach event tags to provide semantic context for cross-regional model collaborative optimization at the central level.

[0117] The above-mentioned technical features form a tight engineering closed loop: the sliding time window structure provides a dynamic data carrier for the weighted empirical distribution function, the weighted empirical distribution function provides a non-parametric comparison benchmark for the calculation of the KS statistic, the calculation result of the KS statistic directly drives the triggering and scope definition of the local incremental training task, and the localized deployment and lightweight upload mechanism jointly ensure the feasibility and real-time performance of the entire closed loop in resource-constrained edge environments. Through the above steps, this invention enables the continuous evolution of a reliable model that supports large-scale distributed new energy power plant clusters with high timeliness, low overhead, and strong privacy protection without uploading raw sensitive environmental data. By employing sliding windows and weighted WEDF, it solves the technical problem of traditional centralized batch updates struggling to track slow environmental drift, improving the model's adaptability to long-term climate trends. By offloading KS tests and QRF training to the local processor, it solves the problems of wide-area communication bandwidth bottlenecks and central computing power overload, significantly shortening the end-to-end latency from data perception to model effectiveness (measured down from hours to minutes). By uploading only parameter increments or statistical summaries, it addresses compliance risks and privacy leaks caused by cross-domain transfer of raw meteorological data, while also reducing the overall operation and maintenance costs of the cloud-edge collaborative architecture. Ultimately, it achieves the self-evolving effect of the reliability assessment model becoming "more accurate and stable with use," providing a reliable online resilience assessment capability for new power systems.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention / invention, and are not intended to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention / invention.

Claims

1. A reliability assessment method for new energy power systems based on environmental variables, comprising acquiring historical and predicted data of environmental variables, power system topology and operation information, generating time-series samples of new energy output, and constructing a two-layer nested reliability assessment architecture, characterized in that: A non-parametric adaptive quantile regression forest model was used to model wind speed and solar radiation intensity, and online incremental updates were triggered based on the Kolmogorov-Smirnov statistic. In the two-layer nested reliability assessment architecture, a cross-scale state coupling interface layer is set up. This interface layer defines a shared state vector that includes the energy storage state of charge, equipment aging index and available backup capacity, and applies Lyapunov stability constraints. A time-based economic value dynamic pricing module based on Q-learning is integrated into the short-term assessment layer, using real-time electricity prices, user interruption contracts, and load elasticity coefficients as the state space to correct load shedding strategies.

2. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, equipment aging index, and available reserve capacity, and uses the shared state vector output by the upper-layer model as the initial boundary condition for the lower-layer evaluation process.

3. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The energy storage state of charge, equipment aging index, and available reserve capacity in the shared state vector are dynamically coupled with external subsystems through a bidirectional power flow interface, an equipment health monitoring system, and a market clearing and demand response platform.

4. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The dual-layer nested reliability assessment architecture includes an upper-layer long-term power transmission system reliability assessment model that operates on an hourly or daily time scale and a lower-layer short-term instantaneous probabilistic reliability assessment process that executes on a minute or second time scale.

5. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The shared state vector defined by the cross-scale state coupling interface layer includes the energy storage state of charge, equipment aging index, and available reserve capacity, and is passed between the upper and lower layers of the two-layer nested architecture through state transition equations.

6. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The cross-scale state coupling interface layer is configured with a state transition error bound constraint module based on the Lyapunov function, which is used to mathematically constrain the deviation between the upper layer output state and the lower layer rolling optimization initial state, and to trigger upper layer model parameter recalibration or lower layer initial condition correction when the state transition error exceeds the stability domain.

7. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The online incremental update mechanism triggers local parameter retraining or the addition of decision tree nodes to the quantile regression forest model when the Kolmogorov-Smirnov statistic exceeds a preset threshold.

8. The reliability assessment method for new energy power systems based on environmental variables according to claim 1, characterized in that: The online incremental update mechanism adopts a sliding time window structure and assigns different confidence weights to historical and current environmental variable data based on a weighted empirical distribution function. The online incremental update deploys the calculation of Kolmogorov-Smirnov statistics and incremental training tasks on a local processor close to the data source, and uploads the model parameter increments or key statistical summaries to the central evaluation server.

Images