Photovoltaic power grid energy storage optimization regulation and control method based on multi-scale prediction

Abstract

The invention relates to a photovoltaic power grid energy storage optimization regulation and control method based on multi-scale prediction. The method comprises the following steps: A1, obtaining historical power generation data, real-time meteorological data and numerical weather forecast of a photovoltaic power station; a2, generating a multi-time-scale photovoltaic output prediction sequence; a3, establishing an energy storage dynamic model of charge and discharge efficiency, capacity attenuation and operation constraint; a4, generating an energy storage charging and discharging demand curve under different time scales; a5, constructing a multi-time scale coupled optimization model by taking power grid operation cost minimization and renewable energy consumption maximization as targets; a6, updating an energy storage scheduling instruction in a rolling manner based on latest prediction data by adopting a model prediction control framework; a7, monitoring the deviation between the actual photovoltaic output and the power grid load, and dynamically adjusting the energy storage charging and discharging power; and A8, correcting a prediction error through Kalman filtering and a closed-loop feedback mechanism. According to the invention, high-efficiency operation can be realized, and power grid cost minimization and renewable energy consumption maximization can be realized.

Description

Technical Field

[0001] This invention relates to the field of energy storage optimization and control technology, and in particular to a photovoltaic power grid energy storage optimization and control method based on multi-scale prediction. Background Technology

[0002] Currently, photovoltaic power generation has become a core force in energy transformation due to its clean and low-carbon characteristics. However, its output is affected by meteorological factors such as sunlight and cloud cover, exhibiting significant intermittency and volatility. Short-term fluctuations can reach ±30% on a second-by-second basis, while medium- and long-term outputs show strong periodic changes with seasons and climate. Large-scale grid connection can easily cause stability problems such as grid frequency deviation and voltage over-limit, resulting in a persistently high curtailment rate.

[0003] While energy storage systems can smooth out fluctuations and improve photovoltaic absorption capacity, traditional energy storage scheduling methods often use a single time scale for optimization, making it difficult to coordinate the contradiction between second-level power balance and seasonal energy transfer. This leads to increasingly prominent problems such as limited photovoltaic absorption rate and shortened energy storage life. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a photovoltaic grid energy storage optimization and control method based on multi-scale prediction to achieve efficient operation.

[0005] To address the aforementioned problems, the present invention provides a photovoltaic power grid energy storage optimization and control method based on multi-scale prediction, characterized by comprising the following steps: A1: Obtain historical power generation data, real-time meteorological data, and numerical weather forecasts for photovoltaic power plants; A2: Construct ultra-short-term, short-term, medium-term, and long-term forecasting models respectively to generate photovoltaic power output forecasting sequences at multiple time scales; A3: Based on the physical characteristics of energy storage devices, establish a dynamic model of energy storage for charging and discharging efficiency, capacity decay, and operational constraints; A4: Combine the energy storage dynamic model obtained in step A3 with the prediction results in step A2 to generate energy storage charging and discharging demand curves at different time scales. A5: To minimize grid operating costs and maximize renewable energy consumption, a multi-time-scale coupled optimization model is constructed. A6: A model predictive control framework is adopted, which includes a rolling optimization problem involving state-space equations, objective functions, and constraints. Energy storage dispatch instructions are updated on a rolling basis based on the latest predictive data. The state-space equations are established based on the energy storage dynamic model, including the SOC evolution equation and the grid power balance equation. The objective function is designed as a weighted combination of multiple time scales, and the specific calculation method is as follows:

[0006] Where: J represents the objective function value; r represents the start time, i.e., the initial moment of optimization; N1 represents the prediction time domain; N2 represents the control time domain; P grid,k Represented as the grid interaction power at time k; SOC k The state of charge (SOC) of the energy storage system at time k is represented by the SOC value. ref P represents the reference state of charge of the energy storage system. curtail,k Let ΔP be the photovoltaic curtailment power at time k. ess,k Let α be the power change rate of the energy storage system at time k; α1, α2, α3, and α4 are weighting factors. A7: Monitor the deviation between actual photovoltaic output and grid load, and dynamically adjust the energy storage charging and discharging power; this includes fast frequency domain compensation, prediction error feedback correction, and dynamic relaxation of safety constraints. A8: Prediction errors are corrected through Kalman filtering and closed-loop feedback mechanisms.

[0007] In step A2, the ultra-short-term prediction refers to using a hybrid model of time series analysis and machine learning, with a time step of 5 minutes. Historical power data, real-time meteorological monitoring data, and component temperature are used as inputs. The ARIMA model is used to capture the short-term autocorrelation characteristics of the time series, while the LSTM neural network is combined to process nonlinear features. An attention mechanism is introduced to dynamically weight the influence of key time points, and the minute-by-minute power prediction sequence for the next 15 to 60 minutes is output.

[0008] In step A2, short-term forecasting refers to physical modeling based on numerical weather forecast data, using a fusion of physical models and data-driven approaches. Theoretical power output is calculated based on solar position, component tilt angle, and irradiance. Satellite cloud image data is then processed through a convolutional neural network to extract spatial features of cloud movement. These features are then fused with the output of the physical model to generate hourly power curves for the subsequent 24-72 hours. Simultaneously, ensemble learning is used to correct system biases.

[0009] In step A2, the medium-term forecast refers to using ensemble forecasting technology to integrate multi-source NWP data, classify similar weather patterns through cluster analysis, and establish a random forest or Gaussian process regression model for each type of weather to output the average daily power output and probability interval for the next 7 days, while also considering seasonal characteristics and introducing periodic features.

[0010] The long-term prediction in step A2 refers to extracting large-scale climate indices, constructing a generalized additive model with historical photovoltaic power output data after dimensionality reduction through principal component analysis, predicting the monthly power generation trend from January to December, and using quantile regression to quantify the probability of the impact of extreme climate events.

[0011] In step A4, when generating energy storage charging and discharging demand curves at different time scales, the ultra-short-term, short-term, medium-term, and long-term photovoltaic power output prediction sequences obtained in step A2 are first time-aligned and normalized to form a standardized power prediction matrix; then a hierarchical decision-making mechanism is adopted for different time scales.

[0012] In step A5, when constructing the multi-timescale coupled optimization model, a system of objective functions containing progressive relationships in the time dimension is first established. At the minute-level time scale, a quadratic penalty is applied to periods when the minute-level fluctuation rate of photovoltaic output exceeds 3% / min, based on the penalty term of frequency domain decomposition. At the hour-level scale, the grid purchase cost, energy storage cycle loss cost, and curtailment penalty cost are uniformly quantified into a cost function. The purchase cost adopts the time-of-use pricing model, the energy storage loss cost is converted into equivalent aging cost through the Rainflow cycle counting method, and the curtailment penalty is inversely proportional to the real-time absorption rate. At the daily and higher scales, the energy storage scheduling priority is dynamically adjusted by calculating the area overlap between the photovoltaic predicted output and the load curve.

[0013] In step A6, within the model predictive control framework, the system collects the latest measured photovoltaic output data, load demand, and energy storage SOC status for each control cycle. Combined with updated ultra-short-term forecast data, the optimization problem is re-solved.

[0014] In step A7, during the fast frequency domain compensation, for second-level fluctuations, a filtering decomposition based on complex wavelet transform is used to extract high-frequency components. An inverse compensation current command is generated using a pre-calibrated energy storage impedance-frequency characteristic curve. The specific calculation method is as follows:

[0015] Among them: I comp (k) indicates a compensation current command; f u Let G(f) be the u-th decomposition frequency; u ) represents the energy storage system at frequency f u The equivalent admittance at a specific frequency is the relationship between the current output capability of energy storage and power fluctuation; Re{·} represents the calculation using the real part of a complex number; W Ψ Represented as Morlet wavelet coefficients; a u This represents the time resolution of the control wavelet transform; ΔP(k,a) u ) represents the original high-frequency power fluctuation signal at time k.

[0016] In step A8, a hybrid estimation strategy is adopted in the prediction error correction stage. Kalman filtering and closed-loop feedback mechanism are deeply integrated to construct a dynamic error compensation system, including a Kalman filtering layer for multi-source data fusion, a closed-loop feedback compensation layer, and a digital twin auxiliary decision-making layer.

[0017] Compared with the prior art, the present invention has the following advantages: 1. This invention achieves efficient operation through multi-source data fusion, multi-scale prediction models, dynamic modeling of energy storage, and optimized scheduling.

[0018] 2. This invention uses an ARIMA+LSTM hybrid algorithm to process minute-level fluctuations, combines a physical model with CNN to analyze cloud image features, and integrates forecasting techniques to address weather uncertainties.

[0019] 3. This invention constructs a battery state-space equation based on electrochemical characteristics to quantify efficiency decay and thermal constraints; then, it uses the MPC framework to continuously optimize charge and discharge commands, combines Kalman filtering to correct prediction errors, and Lyapunov functions to ensure the SOC safety boundary. Attached Figure Description

[0020] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0022] Figure 2 This is a schematic diagram of the multi-scale prediction method of the present invention. Detailed Implementation

[0023] This invention achieves efficient operation through multi-source data fusion, multi-scale prediction models, dynamic modeling of energy storage, and optimized scheduling. First, multi-source data is collected and preprocessed to construct ultra-short-term to long-term multi-scale photovoltaic prediction models, which are then fused to generate continuous sequences. Simultaneously, a dynamic energy storage model is established, including charging and discharging efficiency, capacity decay, and operational constraints, implemented through discretization of state-space equations. Multi-scale energy storage charging and discharging curves are generated based on the prediction results, and a multi-objective optimization model is constructed. Model predictive control is used to continuously update scheduling instructions and dynamically adjust charging and discharging power. Finally, Kalman filtering and closed-loop feedback are used to correct errors, minimizing grid costs and maximizing renewable energy absorption.

[0024] like Figures 1-2 As shown, a photovoltaic power grid energy storage optimization and control method based on multi-scale prediction includes the following steps: A1: Obtain historical power generation data, real-time meteorological data, and numerical weather forecasts for photovoltaic power plants.

[0025] During the data acquisition phase, historical power generation data (including key parameters such as inverter output power, string current and voltage, and module temperature) is first obtained through the SCADA system of the photovoltaic power station, with a time resolution of 1 to 15 minutes, covering the entire annual operating cycle to capture seasonal characteristics. At the same time, real-time monitoring data from meteorological observation stations (including elements such as irradiance, ambient temperature, humidity, and wind speed) are accessed and transmitted in seconds via Modbus / TCP protocol. For numerical weather prediction, gridded data from the GFS global forecast system is used, and ground solar radiation, cloud cover, and precipitation probability forecasts with a time resolution of 1 hour are obtained through the API interface for the next 0 to 240 hours. The gridded data is then matched to the geographical coordinates of the power station using a bilinear interpolation method.

[0026] To improve data quality, outlier removal (based on the 3σ principle), missing value imputation (using time series linear interpolation), and time alignment were performed on the raw data, ultimately forming a multi-source database with a unified timestamp. For power plants with complex terrain, additional drones were deployed for inspection to obtain dust distribution data on component surfaces, or satellite remote sensing surface albedo data was introduced as auxiliary input to correct prediction biases caused by local shading effects.

[0027] A2: Construct ultra-short-term, short-term, medium-term, and long-term forecast models respectively to generate photovoltaic power output forecast sequences at multiple time scales. When constructing multi-scale photovoltaic power output forecast models, differentiated modeling strategies are adopted for different time scales.

[0028] Ultra-short-term forecasting refers to a hybrid model that uses time series analysis and machine learning, with a time step of 5 minutes. It takes historical power data, real-time meteorological monitoring data, and component temperature as inputs, captures the short-term autocorrelation characteristics of the time series through the ARIMA model, and combines the LSTM neural network to process nonlinear features. It also introduces an attention mechanism to dynamically weight the influence of key time points, and outputs a minute-by-minute power forecast sequence for the next 15 to 60 minutes.

[0029] Short-term forecasting refers to a method that combines physical models and data-driven approaches based on numerical weather prediction data. It uses toolkits such as PGLIB for physical modeling, calculates theoretical power output based on solar position, component tilt angle, and irradiance, and then processes satellite cloud image data through convolutional neural networks to extract spatial features of cloud movement. These features are then fused with the output of the physical model to generate hourly power curves for the next 24 to 72 hours. Simultaneously, ensemble learning is used to correct system biases.

[0030] Medium-term forecasting focuses on addressing weather uncertainty by employing ensemble forecasting techniques to integrate multi-source NWP data (such as different initial field forecasts from ECMWF and GFS), classifying similar weather patterns through cluster analysis, and establishing random forest or Gaussian process regression models for each weather category to output the average daily power output and probability range (P10~P90) for the next 7 days, while also considering seasonal characteristics and introducing periodic features.

[0031] Long-term forecasting involves combining climate models with statistical downscaling methods. First, large-scale climate indices are extracted. Then, after dimensionality reduction through principal component analysis, a generalized additive model is constructed with historical photovoltaic power output data to predict the monthly power generation trend from January to December. Quantile regression is used to quantify the probability of the impact of extreme climate events (such as sandstorms and continuous rain).

[0032] The outputs of all models are fused across multiple scales using an adaptive weighting algorithm to generate continuous prediction sequences with temporal resolutions ranging from minutes to months. A rolling training mechanism is employed to ensure the timeliness of the predictions. Specifically, the short-term predictions are at the hourly level, the medium-term at the daily level, and the long-term at the monthly level.

[0033] A3: Based on the physical characteristics of energy storage devices, establish a dynamic model of energy storage for charging and discharging efficiency, capacity decay, and operational constraints.

[0034] When establishing the energy storage dynamic model, an electro-thermal coupling model is first constructed based on the chemical characteristics of the battery. For the electrical characteristics, the Thevenin equivalent circuit model (including parameters such as open-circuit voltage, internal resistance, polarization resistance, and polarization capacitance) is adopted. The relationship between OCV and battery state of charge is obtained through experimental calibration, and the influence of temperature is considered. The charge-discharge efficiency is modeled as a bivariate function of battery state of charge and current rate. The charge-discharge efficiency value is obtained by fitting charge-discharge cycle test data. The specific calculation method is as follows: η(SOC,C) = a·SOC 2 + b·C + c Where: η(SOC,C) represents the charge / discharge efficiency value; SOC represents the battery state of charge; C represents the charge / discharge rate; a, b, and c represent fitting coefficients, constants obtained by fitting experimental data.

[0035] For capacity decay, a semi-empirical model is used to quantify cyclic aging and calendar aging, and the maximum available capacity is dynamically updated. The specific calculation method is as follows:

[0036] Among them: Q max This represents the battery's current maximum usable capacity; Q initialb1 and b2 represent the initial maximum capacity of the battery; b1 and b2 represent empirical fitting coefficients, constants determined through battery cycle test data; N represents the number of charge-discharge cycles the battery has undergone; E a denoted as activation energy, which is the minimum energy required for the aging reaction inside the battery; D represents the ideal gas constant; T represents the absolute temperature of the battery; and t represents the battery's usage time.

[0037] Regarding operational constraints, a SOC operating window is set to prevent overcharging and over-discharging. The charge and discharge power limits take voltage cutoff protection into account, and thermal constraint-triggered derating control is introduced. The specific calculation method for the charge and discharge power limits is as follows: P max = min(P nominal V nominal ×I max ) Where: P max P represents the maximum allowable charge and discharge power of the battery at present. nominal This is expressed as the battery's rated power; V nominal Represented as the battery's cutoff voltage; I max This represents the maximum allowable charge and discharge current of the battery.

[0038] The model is implemented by discretizing the state-space equations, and the specific calculation method is as follows:

[0039] Among them: SOC k+1 Represented as the battery state of charge at time k+1; SOC k Represented by η, the battery state of charge at time k; η represents the charge / discharge efficiency; I k The current is represented as the operating current; Δt represents the time step, i.e., the discretized time interval; Q max,k U represents the maximum available capacity at time k; 1,k Let OCV be the terminal voltage at time k; SOC k R0 represents the equivalent resistance of the instantaneous voltage drop of the battery; V represents the open-circuit voltage. 1,k V is expressed as the polarization voltage at time k; 1,k+1 Let R represent the polarization voltage at time k+1; τ represents the polarization time constant, i.e., the dynamic response speed of the polarization voltage; p It is represented as polarization resistance.

[0040] Simultaneously embeds mixed integer constraints. Forces I / O when in standby mode. k= 0, charging and discharging states are mutually exclusive; model parameters are identified using genetic algorithms based on pulse test and HPPC test data, and updated in real time in BMS; for secondary utilization batteries, an additional single-cell inconsistency model is introduced, and capacity-internal resistance distribution clusters are generated through Monte Carlo simulation to evaluate cluster performance.

[0041] A4: Combine the energy storage dynamic model obtained in step A3 with the prediction results in step A2 to generate energy storage charging and discharging demand curves at different time scales.

[0042] When generating energy storage charging and discharging demand curves at different time scales, the ultra-short-term, short-term, medium-term, and long-term photovoltaic output prediction sequences obtained in step A2 are first time-aligned and normalized to form a standardized power prediction matrix; then a hierarchical decision-making mechanism is adopted for different time scales.

[0043] In the ultra-short-term demand curve, a 15-minute rolling window is used to calculate the difference between the minute-level photovoltaic power forecast and the real-time load curve. When the forecast power exceeds the load demand, a charging command is generated, and when a power deficit occurs, a discharge is triggered. Considering the grid frequency regulation requirements, a regulation dead zone based on frequency deviation (such as no operation within ±0.05Hz) is superimposed on the curve, and a first-order low-pass filter is used to smooth power mutations.

[0044] In the short-term demand curve, based on 24-hour photovoltaic forecasts and time-of-use electricity price signals, dynamic programming is used to solve for the optimal charging and discharging plan. The objective function considers peak-valley price arbitrage and photovoltaic consumption revenue. The constraints include the energy storage SOC continuity equation and grid interconnection power limits. The Lagrange relaxation method is used to handle the coupling constraints and generate hourly charging and discharging power reference values.

[0045] In the medium- and long-term demand curves, historical photovoltaic power output patterns are categorized through cluster analysis (e.g., typical scenarios such as sunny days, cloudy days, and rainy days). Combined with monthly power generation forecasts, a stochastic optimization strategy is used to generate energy storage capacity reservation strategies. For example, when a continuous rainy week is predicted, a State of Charge (SOC) safety margin is set in advance in the demand curve (e.g., maintaining a minimum of 40% energy storage), and the long-term degradation cost of different scheduling strategies is evaluated using a Markov Decision Process (MDP). Finally, a weighted fusion algorithm is used to integrate the curves at various scales. The specific calculation method is as follows: P final (t) = w1·P ultra-short (t) + w2 ·P short (t) ·f smooth (t) Where: P final (t) represents the final combined power curve; P ultra-short (t) represents the ultra-short-term power prediction value; P short(t) represents the short-term power forecast; f smooth (t) represents the smoothing filter function; w1 and w2 represent the weighting factors.

[0046] Meanwhile, the smoothing filter function avoids cross-scale command conflicts, the weighting factor is dynamically adjusted according to the prediction confidence, the demand curve is sent to the energy storage PCS for execution every 5 minutes via the OPC UA protocol, and 10% power bandwidth is reserved to cope with prediction deviations.

[0047] A5: To minimize grid operating costs and maximize renewable energy consumption, a multi-time-scale coupled optimization model is constructed.

[0048] When constructing a multi-timescale coupled optimization model, a system of objective functions containing progressive relationships in the time dimension is first established. At the minute-level timescale, a quadratic penalty is applied to periods when the minute-level fluctuation rate of photovoltaic output exceeds 3% / min, based on the penalty term of frequency domain decomposition. At the hour-level timescale, the grid purchase cost, energy storage cycle loss cost, and curtailment penalty cost are uniformly quantified into a cost function. The purchase cost adopts the time-of-use pricing model, the energy storage loss cost is converted into equivalent aging cost through the Rainflow cycle counting method, and the curtailment penalty is inversely proportional to the real-time absorption rate. At the daily and higher-level timescales, the energy storage scheduling priority is dynamically adjusted by calculating the area overlap between the photovoltaic predicted output and the load curve.

[0049] To achieve multi-scale target coupling, a two-layer optimization architecture is adopted: the upper layer solves the energy storage SOC plan curve on a daily basis, and transforms the long-term target into the boundary conditions for hourly optimization through Benders decomposition; the lower layer performs real-time optimization at 15-minute intervals and feeds back the actual operation data to the upper layer to correct prediction errors; the model solution adopts an improved mixed integer second-order cone programming algorithm to perform piecewise linearization of the non-convex energy storage efficiency constraints.

[0050] A6: A model predictive control framework is adopted, which includes a rolling optimization problem involving state-space equations, objective functions, and constraints, and energy storage scheduling instructions are updated on a rolling basis based on the latest prediction data.

[0051] In the model predictive control framework, for each control cycle (e.g., 5 minutes), the system collects the latest measured photovoltaic output data, load demand, and energy storage SOC status, and combines this with updated ultra-short-term forecast data (for the next 15-30 minutes) to resolve the optimization problem.

[0052] The state-space equations are established based on the energy storage dynamic model, including the SOC evolution equation and the grid power balance equation.

[0053] The objective function is designed as a weighted combination of multiple time scales, and the specific calculation method is as follows:

[0054] Where: J represents the objective function value; r represents the start time, i.e., the initial moment of optimization; N1 represents the prediction time domain; N2 represents the control time domain; P grid,k Represented as the grid interaction power at time k; SOC k The state of charge (SOC) of the energy storage system at time k is represented by the SOC value. ref P represents the reference state of charge of the energy storage system. curtail,k Let ΔP be the photovoltaic curtailment power at time k. ess,k α1 represents the power change rate of the energy storage system at time k; α2, α3, and α4 represent weighting factors.

[0055] During the rolling optimization process, a dynamic weight adjustment strategy is adopted. When the photovoltaic prediction error increases (such as when the cloud moves rapidly), the α1 weight is automatically increased to enhance the power grid smoothing effect. When the energy storage SOC approaches the boundary, the α2 weight is increased to force the SOC to return to the safe range.

[0056] Robust optimization is embedded in MPC to generate interval predictions of photovoltaic output based on the statistical distribution of historical prediction errors, and a two-stage robust optimization model is constructed. The first stage makes decisions on the energy storage ground state dispatch instructions, and the second stage adjusts the reserve power to cope with the worst prediction deviation scenario. The final output dispatch instructions are sent to the energy storage converter after low-pass filtering, and the status information is synchronized with the grid energy management system through the OPC UA protocol.

[0057] A7: Monitors the deviation between actual photovoltaic output and grid load, and dynamically adjusts the energy storage charging and discharging power. It continuously acquires the actual output of the photovoltaic array through a high-speed data acquisition module and calculates the instantaneous power deviation. This deviation signal undergoes a three-level processing flow to achieve dynamic correction of the energy storage power, including fast frequency domain compensation, prediction error feedback correction, and dynamic relaxation of safety constraints.

[0058] In fast frequency domain compensation, for second-level fluctuations (0.1~2Hz frequency band), a filtering decomposition based on complex wavelet transform is used to extract high-frequency components. Using a pre-calibrated energy storage impedance-frequency characteristic curve, an inverse compensation current command is generated. The specific calculation method is as follows:

[0059] Among them: I comp (k) indicates a compensation current command; f u Let G(f) be the u-th decomposition frequency; u ) represents the energy storage system at frequency f u The equivalent admittance at a specific frequency is the relationship between the current output capability of energy storage and power fluctuation; Re{·} represents the calculation using the real part of a complex number; W ΨRepresented as Morlet wavelet coefficients; a u This represents the time resolution of the control wavelet transform; ΔP(k,a) u ) represents the original high-frequency power fluctuation signal at time k, N represents the number of decomposed frequency layers, and u represents the number of decomposed frequency layers.

[0060] In the prediction error feedback correction, the average deviation is calculated every 5 minutes. If it exceeds the threshold, the MPC model is re-optimized. The prediction error model parameters are updated using a recursive least squares method with a forgetting factor. The specific calculation method is as follows:

[0061] Where: θ k+1 θ is represented as the updated model parameter vector at time k+1. k U is represented as the updated model parameter vector at time k; k The covariance matrix at time k is denoted as ф. k Let be the meteorological feature vector at time k; λ represents the forgetting factor. Represented as meteorological feature vector ф k transpose of; P k Let y be the parameter estimation error covariance matrix at time k; k This is expressed as the actual output value of photovoltaic power.

[0062] In dynamic relaxation of safety constraints, when a sustained large deviation leads to the risk of exceeding the SOC limit, emergency control based on the Lyapunov energy function is initiated, and a power support request is sent to the power grid dispatch center, triggering coordinated control of multiple energy storage units. The specific method for calculating the Lyapunov energy function value is as follows:

[0063] Where: V(SOC) represents the Lyapunov energy function value; SOC represents the battery state of charge; SOC safe This is represented as the safe operating threshold.

[0064] All adjustment commands are verified by CRC and transmitted to the energy storage PCS via the CAN bus. Three-level PWM modulation is used to achieve a power tracking accuracy of ±0.5%. The system synchronously generates a deviation analysis report, including: photovoltaic prediction error spectrum distribution map; energy storage compensation energy statistics; SOC trajectory deviation assessment, used for offline optimization of prediction model parameters; under extreme weather conditions, it automatically switches to robust tracking mode based on Kalman filtering, treating photovoltaic output as an unmeasurable disturbance and allocating energy storage power only according to load deviation.

[0065] A8: Prediction errors are corrected through Kalman filtering and closed-loop feedback mechanisms.

[0066] In the prediction error correction stage, a hybrid estimation strategy is adopted, which deeply integrates Kalman filtering with closed-loop feedback mechanism to construct a dynamic error compensation system, including a Kalman filtering layer for multi-source data fusion, a closed-loop feedback compensation layer, and a digital twin-assisted decision-making layer.

[0067] In the Kalman filter layer of multi-source data fusion, the photovoltaic output error propagation equation is established based on the state-space model. The observation vector is constructed by real-time acquisition of inverter output current, string voltage and irradiance data from the SCADA system. An adaptive Kalman gain algorithm is adopted to automatically increase the process noise covariance matrix when rapid cloud movement is detected, thereby enhancing the filter's ability to track abrupt changes. The error estimate is updated every 30 seconds, and numerical stability is ensured through Cholesky decomposition.

[0068] In the closed-loop feedback compensation layer, the PID controller output is limited and then directly corrects the current energy storage scheduling command. It is then injected back into the prediction model input layer to form a feedforward-feedback composite control structure. When a continuous unidirectional deviation is detected, the learning mechanism is automatically activated, and the PID parameters are dynamically adjusted through the online gradient descent method.

[0069] In the digital twin-assisted decision-making layer, a real-time simulation image of the photovoltaic power station is constructed, and the actual operating data is spatiotemporally aligned and compared with the simulation prediction results. An error pattern recognition technology based on dynamic time warping algorithm is adopted to automatically classify the deviation type and match it with a preset compensation strategy library.

[0070] The system is configured with a three-level error handling mechanism: when |ΔZ|<5%, only logs are recorded; when 5%≤|ΔZ|<15%, Kalman filter correction is triggered; when |ΔZ|≥15%, a full closed-loop reconstruction mode is initiated, including: pausing the current prediction model, switching to nonparametric prediction based on similar day retrieval, and sending an early warning signal to the dispatch center.

[0071] All correction operations are recorded on the blockchain to ensure the traceability of the decision-making process; zero-value drift calibration is automatically performed at night or in the absence of light to eliminate accumulated sensor errors.

[0072] In summary, this invention constructs a four-layer prediction model (LSTM + physical fusion + ensemble forecast + climate downscaling) ranging from minute to month, combines it with an energy storage dynamic model that couples electricity, heat, and aging, and adopts a two-layer optimization and adaptive control strategy to smooth photovoltaic fluctuations and improve the absorption rate. Furthermore, by incorporating attention mechanisms, Lyapunov emergency control, and other technologies, it solves the problem of stable operation of high-proportion renewable energy grids.

Claims

1. A photovoltaic power grid energy storage optimization and control method based on multi-scale prediction, characterized in that, Includes the following steps: A1: Obtain historical power generation data, real-time meteorological data, and numerical weather forecasts for photovoltaic power plants; A2: Construct ultra-short-term, short-term, medium-term, and long-term forecasting models respectively to generate photovoltaic power output forecasting sequences at multiple time scales; A3: Based on the physical characteristics of energy storage devices, establish a dynamic model of energy storage for charging and discharging efficiency, capacity decay, and operational constraints; A4: Combine the energy storage dynamic model obtained in step A3 with the prediction results in step A2 to generate energy storage charging and discharging demand curves at different time scales. A5: To minimize grid operating costs and maximize renewable energy consumption, a multi-time-scale coupled optimization model is constructed. A6: A model predictive control framework is adopted, which includes a rolling optimization problem involving state-space equations, objective functions, and constraints. Energy storage dispatch instructions are updated on a rolling basis based on the latest predictive data. The state-space equations are established based on the energy storage dynamic model, including the SOC evolution equation and the grid power balance equation. The objective function is designed as a weighted combination of multiple time scales, and the specific calculation method is as follows: ， Where: J represents the objective function value; r represents the start time, i.e., the initial moment of optimization; N1 represents the prediction time domain; N2 represents the control time domain; P grid,k Represented as the grid interaction power at time k; SOC k The state of charge (SOC) of the energy storage system at time k is represented by the SOC value. ref P represents the reference state of charge of the energy storage system. curtail,k Let ΔP be the photovoltaic curtailment power at time k. ess,k α1, α2, α3, and α4 represent the power change rate of the energy storage system at time k; k represents the control time number. A7: Monitor the deviation between actual photovoltaic output and grid load, and dynamically adjust the energy storage charging and discharging power; this includes fast frequency domain compensation, prediction error feedback correction, and dynamic relaxation of safety constraints. A8: Prediction errors are corrected through Kalman filtering and closed-loop feedback mechanisms.

2. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A2, the ultra-short-term prediction refers to using a hybrid model of time series analysis and machine learning, with a time step of 5 minutes. Historical power data, real-time meteorological monitoring data, and component temperature are used as inputs. The ARIMA model is used to capture the short-term autocorrelation characteristics of the time series, while the LSTM neural network is combined to process nonlinear features. An attention mechanism is introduced to dynamically weight the influence of key time points, and the minute-by-minute power prediction sequence for the next 15 to 60 minutes is output.

3. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A2, short-term forecasting refers to physical modeling based on numerical weather forecast data, using a fusion of physical models and data-driven approaches. Theoretical power output is calculated based on solar position, component tilt angle, and irradiance. Satellite cloud image data is then processed through a convolutional neural network to extract spatial features of cloud movement. These features are then fused with the output of the physical model to generate hourly power curves for the subsequent 24-72 hours. Simultaneously, ensemble learning is used to correct system biases.

4. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A2, the medium-term forecast refers to using ensemble forecasting technology to integrate multi-source NWP data, classify similar weather patterns through cluster analysis, and establish a random forest or Gaussian process regression model for each type of weather to output the average daily power output and probability interval for the next 7 days, while also considering seasonal characteristics and introducing periodic features.

5. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: The long-term prediction in step A2 refers to extracting large-scale climate indices, constructing a generalized additive model with historical photovoltaic power output data after dimensionality reduction through principal component analysis, predicting the monthly power generation trend from January to December, and using quantile regression to quantify the probability of the impact of extreme climate events.

6. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A4, when generating energy storage charging and discharging demand curves at different time scales, the ultra-short-term, short-term, medium-term, and long-term photovoltaic power output prediction sequences obtained in step A2 are first time-aligned and normalized to form a standardized power prediction matrix; then a hierarchical decision-making mechanism is adopted for different time scales.

7. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A5, when constructing the multi-timescale coupled optimization model, a system of objective functions containing progressive relationships in the time dimension is first established. At the minute-level time scale, a quadratic penalty is applied to periods when the minute-level fluctuation rate of photovoltaic output exceeds 3% / min, based on the penalty term of frequency domain decomposition. At the hour-level scale, the grid purchase cost, energy storage cycle loss cost, and curtailment penalty cost are uniformly quantified into a cost function. The purchase cost adopts the time-of-use pricing model, the energy storage loss cost is converted into equivalent aging cost through the Rainflow cycle counting method, and the curtailment penalty is inversely proportional to the real-time absorption rate. At the daily and higher scales, the energy storage scheduling priority is dynamically adjusted by calculating the area overlap between the photovoltaic predicted output and the load curve.

8. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A6, within the model predictive control framework, the system collects the latest measured photovoltaic output data, load demand, and energy storage SOC status for each control cycle. Combined with updated ultra-short-term forecast data, the optimization problem is re-solved.

9. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A7, during the fast frequency domain compensation, for second-level fluctuations, a filtering decomposition based on complex wavelet transform is used to extract high-frequency components. An inverse compensation current command is generated using a pre-calibrated energy storage impedance-frequency characteristic curve. The specific calculation method is as follows: ， Among them: I comp (k) indicates a compensation current command; f u Let G(f) be the u-th decomposition frequency; u ) represents the energy storage system at frequency f u The equivalent admittance at a specific frequency is the relationship between the current output capability of energy storage and power fluctuation; Re{·} represents the calculation using the real part of a complex number; W Ψ Represented as Morlet wavelet coefficients; a u This represents the time resolution of the control wavelet transform; ΔP(k,a) u ) represents the original high-frequency power fluctuation signal at time k, N represents the number of decomposed frequency layers, and u represents the number of decomposed frequency layers.

10. The photovoltaic grid energy storage optimization and control method based on multi-scale prediction as described in claim 1, characterized in that: In step A8, a hybrid estimation strategy is adopted in the prediction error correction stage. Kalman filtering and closed-loop feedback mechanism are deeply integrated to construct a dynamic error compensation system, including a Kalman filtering layer for multi-source data fusion, a closed-loop feedback compensation layer, and a digital twin auxiliary decision-making layer.

Images