Energy storage system optimal configuration method based on output deviation real-time adaptive compensation

Abstract

The invention discloses an energy storage system optimal configuration method based on output deviation real-time adaptive compensation, and particularly relates to the field of new energy power generation and energy storage optimal configuration. According to the method, a similarity structure analysis algorithm and an opportunity compensation cost mechanism are introduced, photovoltaic power deviation characteristics are dynamically extracted, and a deviation target domain is established; and solving by combining a multi-target Harris Hawk algorithm (HHO) to obtain a Pareto optimal configuration scheme of the capacity and quantity of the energy storage system, and realizing adaptive optimization and output deviation minimization of the energy storage system. According to the method, the structural mode of the photovoltaic output deviation is extracted by using a similarity structural analysis algorithm, a structured recognition approach of the deviation mode is established instead of depending on mean fitting or curve regression, differential compensation reference values can be output for different meteorological states, and energy storage control is changed from fluctuation tracking to deviation feedforward compensation.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and energy storage optimization configuration technology, and more specifically, to an energy storage system optimization configuration method based on real-time adaptive compensation for output deviation. Background Technology

[0002] With the advancement of my country's "dual-carbon" strategic goals, the integration of a high proportion of renewable energy into the power system has become a trend. Photovoltaic power plants are rapidly developing in high-altitude areas due to their clean and sustainable characteristics. However, the sunlight conditions in high-altitude regions are affected by multiple factors such as terrain, climate, snow cover, and cloud reflection, resulting in strong randomness and non-stationarity in photovoltaic power output, leading to frequent fluctuations in its power curve. The integration of large amounts of fluctuating photovoltaic output into the grid causes rapid changes in net load, thereby affecting the voltage stability and frequency security of the power grid.

[0003] Traditional energy storage configuration methods are typically optimized based on a single economic indicator (such as Levelized Cost of Energy (LCOE) or Return on Investment (ROI), ignoring the dynamic changes in photovoltaic power output in high-altitude environments. This static optimization method struggles to compensate for photovoltaic power deviations in real time, leading to problems such as capacity redundancy, response lag, or excessively high idle rates in the energy storage system during operation, thereby reducing the overall economic efficiency and operational reliability of the system.

[0004] Therefore, there is an urgent need for an energy storage optimization configuration method that can perform real-time adaptive analysis and dynamic compensation of photovoltaic output deviation under complex climatic conditions, so as to achieve an energy storage system design that balances economy and reliability. Summary of the Invention

[0005] To address the issues of severe fluctuations and significant prediction errors in photovoltaic (PV) output under complex high-altitude climate conditions, a method for optimizing energy storage system configuration based on real-time adaptive compensation for output deviation is proposed. This method dynamically extracts PV power deviation characteristics by introducing a similarity structure analysis algorithm and an opportunity compensation cost mechanism, establishes a deviation target domain, and solves it using the multi-objective Harris Hawk (HHO) algorithm to obtain the Pareto optimal configuration scheme for the capacity and number of energy storage systems, achieving adaptive optimization of the energy storage system and minimizing output deviation.

[0006] Furthermore, the above method consists of the following steps: Step 1: Acquisition and Preprocessing of High-Altitude Solar Power Output Data (1) Data acquisition In response to the challenges of complex climate, difficult meteorological data sampling, and significant impact of terrain-induced shading on sunlight in high-altitude regions, this invention proposes a multi-source fusion method for acquiring high-altitude photovoltaic power output data, comprising the following: ① Real-time operation data collection of photovoltaic power plants The SCADA system collects real-time information on photovoltaic module string voltage, current, power, temperature, and combiner box status; the sampling period is set to 5 minutes and stored in time series format for short-term dynamic output analysis; the inverter output active power and bus voltage and current data are collected simultaneously to form a high-resolution photovoltaic output record.

[0007] ②Simultaneous acquisition of meteorological and geographic data Meteorological parameters such as solar irradiance, wind speed, wind direction, temperature, humidity, air pressure, snow thickness, and cloud cover are obtained through on-site meteorological stations. Digital elevation models (DEM) and terrain shading correction algorithms are used to calculate correction parameters for the angle and intensity of terrain shading.

[0008] ③ Integration of historical output and forecast data The system integrates long-term sunshine forecast data from meteorological departments with historical processed data; it uses time-synchronous interpolation to complete missing historical samples; and it constructs a unified photovoltaic power-meteorological parameter database to provide data support for subsequent deviation analysis and modeling.

[0009] (2) Data preprocessing To ensure the stability and analyzability of photovoltaic power output data, this invention employs a multi-level data cleaning and feature decomposition mechanism, with the following steps: ①Outlier detection and cleaning Outliers in the photovoltaic power sequence are identified using the three-σ criterion and box plot detection method; abnormal segments caused by sensor failure, inverter shutdown, and snow shading are marked; short-term outliers are corrected using the sliding window mid-range filtering method; and continuous missing measurement segments are reconstructed and completed using ARIMA.

[0010] ② Noise reduction and trend decomposition Wavelet packet decomposition (WPD) was used to perform multi-scale decomposition of the photovoltaic power series, separating the long-term trend term from the high-frequency noise term: the high-frequency component corresponds to random disturbances and measurement noise; the low-frequency component corresponds to climate change trends and solar cycle characteristics. Only the low-frequency subband signal was retained for long-term power output trend analysis to improve the model's stability.

[0011] ③ Normalization and Reconfiguration of Photovoltaic Output Characteristics Each characteristic parameter (irradiance, temperature, wind speed, etc.) is subjected to minimum-maximum normalization, with a range of 0 to 1; and a terrain correction factor is applied. The weighted correction to the normalization result is calculated using the following formula: , in, It is calculated from topographic factors such as slope, aspect, and shading angle.

[0012] ④ Dataset partitioning and feature annotation Based on the principle of temporal continuity, the data was divided into a training set (70%), a validation set (15%), and a test set (15%). Each sample data was labeled with its timestamp, corresponding meteorological state label (sunny, cloudy, snowy, foggy, etc.), and actual power deviation, providing an input basis for subsequent adaptive deviation analysis.

[0013] (3) Data feature extraction and modeling interface After preprocessing, feature extraction is performed on the cleaned photovoltaic output data, mainly including: Intraday fluctuation range of photovoltaic power output Power change rate Smoothness index Coupling coefficient between light and temperature Output periodicity index .

[0014] The extracted features are passed as input variables to the next step, the "Photovoltaic Output Deviation Adaptive Extraction Algorithm" module, for training the similarity structure analysis model and constructing the deviation target domain.

[0015] Step 2: Adaptive Extraction of Photovoltaic Output Deviation (1) Overall Approach Photovoltaic power output in high-altitude areas is affected by complex climate and terrain conditions, and the deviation between predicted and actual power output exhibits nonlinear, non-stationary, and multi-scale coupled characteristics. Traditional deviation modeling methods, such as linear regression and simple time moving average, cannot effectively capture the dynamic structural features of power output deviation. To address this, this invention proposes an adaptive extraction algorithm for photovoltaic power output deviation based on similarity structure analysis (SSA). This algorithm constructs a multidimensional similarity matrix between predicted and historical power output, dynamically identifies the statistical structural features of power deviation, and updates the deviation target domain in real time through an adaptive threshold mechanism, achieving accurate extraction and self-learning correction of photovoltaic power output deviation.

[0016] (2) Algorithm principle Assume the predicted photovoltaic output value at time t is The actual observed output is Its power deviation is defined as: , Deviation sequence under complex climatic conditions It exhibits multiple temporal correlations. This algorithm performs adaptive extraction through the following three core processes: ① Similarity structure modeling: Analyze the multidimensional similarity between predicted output and historical output, and extract the deviation features of similar time periods.

[0017] ② Deviation pattern recognition: Through clustering and structure mapping, the deviation distribution patterns under different meteorological conditions are identified.

[0018] ③ Adaptive update of target domain: The deviation target domain is dynamically adjusted according to the real-time deviation trend to realize the self-learning feature of the model.

[0019] (3) Algorithm implementation process ① Construct a multidimensional feature matrix Transform photovoltaic power output data into a multi-dimensional input vector: , in, Solar irradiance; :temperature; Wind speed; :humidity; : Incident angle correction factor; : Predicted output value.

[0020] At the same time, define the corresponding actual processing vector: , Constructing the feature set: , Used for subsequent similarity calculations.

[0021] ② Construction of similarity structure matrix The Gaussian kernel similarity function is used to measure the similarity between any two time points. Processing structural similarity: , in, The kernel width parameter is adaptively set based on the data standard deviation.

[0022] To improve the temporal sensitivity of similarity, a time decay weight is introduced: , Construct a weighted similarity matrix: , This matrix reflects the dynamic structural correlation of photovoltaic power output under different meteorological and temporal conditions.

[0023] ③ Biased clustering and structure mapping Through the Perform spectral clustering to divide all samples into A cluster of deviation patterns: , Each cluster represents a typical output deviation characteristic structure (e.g., high irradiation-low temperature region, low irradiation-high humidity region, etc.).

[0024] Within each cluster, calculate the corresponding mean deviation eigenvector: , Further establish the structural mapping relationship between "meteorological conditions and deviation models": , This mapping function is used to identify the power output deviation model under the current meteorological conditions in real time.

[0025] ④ Adaptive determination of the deviation target domain Define the target domain of deviation as: , To achieve real-time adaptive updates, this invention introduces a sliding time window mechanism (window length L) and dynamic threshold update rules: , in It is an adaptive learning rate.

[0026] This mechanism can quickly adjust the upper and lower limits of the deviation range under sudden meteorological changes (such as snow accumulation or cloud shadows), achieving dynamic self-learning.

[0027] ⑤ Deviation feature extraction and compensation signal output Finally, the real-time deviation compensation signal is calculated: , in, It is a real-time power compensation command for the energy storage system. It is a compensation coefficient that is dynamically adjusted based on historical performance.

[0028] The output results serve as input to the subsequent dynamic compensation model of the energy storage system, guiding the energy storage system to perform adaptive charge and discharge control.

[0029] Step 3: Establishment of Dynamic Compensation Model for Energy Storage System (1) Energy state modeling of energy storage system The dynamic characteristics of an energy storage system can be characterized by its state of energy (SOC). The rated capacity of the radio energy storage system is... E max， exist t The SOC at time t is defined as: , in E ( t The current energy level of the energy storage unit is denoted as SOC. The change in SOC over time is determined by the power input and output. , in, and These represent the discharge and charging power, respectively. These represent the discharge and charge efficiencies, respectively. This model allows for real-time calculation of the energy storage system's energy changes, ensuring it remains within the permissible State of Charge (SOC) range. Internal operation.

[0030] (2) Power dynamic response model of energy storage system To describe the response characteristics of an energy storage system to fluctuations in photovoltaic power output, a power dynamic model needs to be established. The energy storage system output power... Driven by the target power deviation signal and affected by system response delay and dynamic constraints: , in, This refers to the photovoltaic power output deviation obtained through the deviation extraction algorithm; This is the system gain coefficient, used to control the compensation strength; The time constant representing the power response of the energy storage system reflects the system's inertial characteristics. By discretizing this equation, dynamic simulation and control strategy solutions based on real-time data can be achieved.

[0031] (3) Design of dynamic compensation strategy Based on the deviation detection and power response model, a dynamic compensation strategy for the energy storage system is designed. The core idea is to adjust the energy storage power output in real time according to the amplitude and trend of the photovoltaic power output deviation, thereby achieving energy balance and output smoothing. Its control strategies can be divided into the following three categories: ① Fast power compensation tracking: When photovoltaic power output fluctuates rapidly (such as sudden changes in cloud shadows), the energy storage system uses the power error as input and employs proportional-integral control: , This strategy can quickly suppress short-term power fluctuations and maintain output continuity.

[0032] ② Slow time domain energy balance compensation: For continuous irradiance variations or diurnal power imbalances, energy allocation is based on the SOC regulation mechanism: , By gradually adjusting the charging and discharging power, the energy storage system can be maintained to operate within a safe energy range.

[0033] ③ Adaptive compensation optimization: Introducing a fuzzy logic mechanism to dynamically adjust based on historical deviation distribution and system operating status. These parameters enable adaptive energy regulation across multiple time scales.

[0034] ④ Model Implementation and Simulation Verification The established energy storage dynamic compensation model is embedded into the power balance equation of the high-altitude photovoltaic system, resulting in: , The model was simulated and validated using the MATLAB platform, which evaluated the impact of different compensation parameters and control strategies on the system output stability. The results show that the model can effectively reduce the fluctuation amplitude of high-altitude photovoltaic output and improve the system power smoothing coefficient and grid connection quality.

[0035] Step 4: Construction of Multi-Objective Optimization Model (1) Overall Model Idea The power output of high-altitude photovoltaic systems is significantly affected by complex climatic environments, exhibiting substantial fluctuations and uncertainties. Dynamic compensation through energy storage systems can improve power smoothness, but energy storage configurations under different capacities and control strategies can significantly impact system economics. Therefore, this model aims to find the optimal balance between economics and stability, achieving the joint determination of optimal energy storage capacity, power ratio, and operating strategy through a multi-objective optimization method.

[0036] (2) Optimize the design of the objective function Objective 1: Minimize the annual overall cost of the energy storage system Annual comprehensive cost of energy storage system Including investment depreciation costs, operation and maintenance costs, and energy loss costs over the lifespan, it can be expressed as: , in: Initial investment cost of energy storage equipment; Investment cost per unit capacity and power; Energy storage system lifespan (years); Operating and maintenance costs; Energy loss costs caused by energy storage round-trip efficiency; Round-trip power efficiency.

[0037] In addition, the levelized cost of electricity (LCOE) for energy storage is introduced as a long-term economic indicator: , in This represents the total discharge capacity of the energy storage system over its lifespan.

[0038] One of the optimization goals is to make Minimum.

[0039] Objective 2: Minimize photovoltaic power output deviation rate The deviation rate of photovoltaic output measures the system's smoothing ability and grid friendliness, and is defined as: , in: Combined output power of photovoltaic and energy storage systems; : Expected output power curve; N : Number of time steps.

[0040] To reflect the effectiveness of the system's operation, the following auxiliary performance indicators are also introduced: Energy storage system idle rate: , Reflects the utilization rate of energy storage devices; Energy storage load tracking error: , The final objective function in its synthesized form is: , in These are weighting coefficients used to balance various performance indicators.

[0041] (3) Constraints The optimization model must meet the physical and technical constraints of energy storage and system operation: ① Power constraint: , ② Energy constraint (SOC): , And satisfy energy balance: , ③System power balance: , ④ Service life constraints: The number of energy storage cycles shall not exceed the maximum allowable value: , Step 5: Solve using the multi-objective Harris Eagle optimization algorithm (1) Overall framework of the algorithm The Harris Eagle Optimization Algorithm (HHO) is a biomimetic swarm intelligence optimization algorithm that simulates the hunting behavior of falcon flocks for global search and local exploitation. To extend HHO into a multi-objective solver, this invention introduces the following improvements to the standard HHO framework: ① Dynamic Learning Factor (DLF) — Enables individuals to adaptively adjust step size and exploration / utilization balance during the iteration process to cope with the non-stationarity of photovoltaic power output deviation.

[0042] ② Non-dominated ranking strategy—used to rank candidate solutions hierarchically according to the merits of multiple objectives, and to maintain the diversity of solutions by combining crowding or distance indicators.

[0043] ③ Pareto boundary dynamic update mechanism—Introducing external archives to save the currently found Pareto solutions and dynamically cleaning and updating them to ensure the maintenance and evolution of high-quality solution sets.

[0044] (2) Decoding and feasible region Decision variable vector x: Represents variables related to energy storage configuration and operation, such as the rated energy of each energy storage unit. E i Rated power P i Number of units n i Lifetime options, backup capacity distribution, etc. If the system consists of... M If the energy storage is composed of various groups, then: , Variable range: Define upper and lower bounds for each variable. , (3) Objective function and constraint evaluation Given a decision vector x, compute the bi-objective. Combining constraints, an optimized fitness vector (for non-dominated ranking) is formed through a penalty mechanism: , Where g(x) represents the merging quantity of all inequality constraint violations. m A relatively large penalty coefficient (dynamically adjustable).

[0045] (4) Detailed design of key improvement points ① Dynamic Learning Factor (DLF) DLF is used to control the search stride and adaptive learning rate of each generation of falcons (individuals), and is defined as: , in t For the current iteration, T max The maximum number of iterations, k,g To adjust the parameters, DLF is kept at a large value in the early stages to enhance exploration, and then gradually reduced in the later stages to improve development accuracy; the noise term helps to cross local optima.

[0046] ② DLF integration in HHO action strategy In HHO's different hunting strategies (encirclement, ambush, encirclement, rapid pursuit, etc.), the step size in the individual update formula is multiplied by... ,For example: , ③ Non-dominated sorting + crowding (or distance) maintenance Perform a non-dominated sort on the solution set after merging the current population with external archives, and then stratify the solutions into fronts. F 1, F 2,…).

[0047] Within the same frontier, solutions are ranked using crowding distance or a homogeneity metric based on the target space to select candidates that maintain diversity for the next generation or archive.

[0048] Selection strategy: Prioritize retaining the more advanced non-dominated layers, and within each layer, select from the most crowded to the least crowded, until the population / archive capacity is full.

[0049] ④ Pareto boundary dynamic update mechanism Maintain a finite-capacity external archive A to store currently discovered non-dominated solutions. Each time a new solution is generated: merge the new solution with the solutions in A; delete dominated solutions; if the archive size exceeds the limit, remove redundant solutions based on crowding or target space uniformity to ensure Pareto boundary coverage and diversity; periodically perform a mutation to introduce slight variations into the archive to expand boundary exploration.

[0050] Archive updates use a dynamic threshold: when the improvement in the objective function value is less than a threshold... ϵ At the same time, the standards for archival retention should be tightened to improve the quality of data collection.

[0051] ⑤ Constraint handling strategy Main strategy: Repair + Penalty. For easily repairable constraints (such as integerization and power upper bounds), a repair function is used to directly map them to the feasible region; for difficult-to-repair or complexly coupled constraints, an adaptive penalty is employed. , The penalty is adaptively adjusted according to the proportion of constraints not met, prompting the search to gradually converge toward the feasible region.

[0052] Step 6: Adaptive Implementation and Online Correction (1) Real-time monitoring and data acquisition module The system deploys multi-source data acquisition terminals to monitor in real time the photovoltaic array output, grid-side power, energy storage SOC status, meteorological parameters (irradiance, temperature, wind speed, etc.) and load-side power demand of the high-altitude photovoltaic power station.

[0053] The sampled data undergoes preliminary noise reduction and outlier detection at edge computing nodes. Outliers are removed using moving average filtering and the triple σ criterion, ensuring the timeliness and reliability of the input data. The data acquisition cycle is adaptively adjusted according to the operating scenario, typically set to 1-5 minutes to balance monitoring accuracy and communication bandwidth.

[0054] (2) Dynamic update mechanism for the deviation target domain Using the photovoltaic power output deviation adaptive extraction algorithm established in step 2, the system continuously calculates the deviation distribution sequence between the predicted photovoltaic power output and the measured power output during operation.

[0055] Based on a sliding time window mechanism (e.g., the past 24 hours), similarity structure analysis (SSA) is used to re-identify the photovoltaic power output mode. The latest deviation statistical characteristics are input into the deviation target domain to update the model, and the power deviation boundary conditions are dynamically corrected. When the deviation target domain changes significantly, the system automatically triggers the adjustment of the energy storage scheduling strategy.

[0056] (3) Adaptive scheduling strategy for energy storage charging and discharging Based on the updated deviation target domain and real-time load demand, the system employs an adaptive scheduling strategy to dynamically control the charging and discharging power of the energy storage system. Specifically, this includes: Charging strategy: When the photovoltaic output is higher than the predicted value and the SOC is lower than the safe limit, the system prioritizes fast charging to absorb the excess energy; Discharge strategy: When the photovoltaic output is lower than the predicted value or the net load of the grid is higher than the target value, the system dynamically releases energy according to the deviation target domain; Smooth control: Introducing proportional-integral-derivative (PID) control combined with fuzzy logic to suppress frequent start-stop of energy storage and extend equipment life.

[0057] The system adopts an economic constraint based on opportunity compensation cost to conduct real-time assessment of energy loss and depreciation caused by frequent energy storage operations, thereby achieving dual optimization of economy and stability.

[0058] (4) Model parameter self-learning correction mechanism To enhance long-term adaptive capabilities, the system incorporates a lightweight self-learning module to periodically correct key model parameters (such as bias mapping coefficients, SOC dynamic thresholds, and opportunity cost weights) online. The recursive least squares (RLS) method and an improved particle filter (PF) algorithm are used to jointly update parameters, ensuring that the model maintains high-precision prediction and control performance even under abrupt changes in climate conditions (such as snow accumulation, strong winds, or low-temperature radiation attenuation).

[0059] (5) Closed-loop optimization and execution feedback After each operating cycle, the system feeds back the results, such as the energy storage operating status, photovoltaic deviation compensation effect, and economic indicators (e.g., levelized cost of equity, LCOE), to the optimization model module, forming a data closed loop. Through the periodically triggered MO-HHO algorithm re-optimization process, the system can quickly adjust the Pareto solution set without re-initialization, realizing the transformation from "static optimization" to "dynamic self-evolutionary optimization".

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problem that photovoltaic (PV) output exhibits strong randomness and non-stationarity under complex meteorological conditions in high-altitude areas, and that traditional energy storage capacity configuration based solely on static economic indicators is insufficient for effective deviation compensation. By constructing an adaptive identification method for PV output deviation structures and introducing an online dynamic correction mechanism, the energy storage system no longer relies on fixed capacity configurations but can achieve real-time compensation based on the deviation target domain, thereby significantly improving the energy storage system's adaptability to PV output deviations. This invention utilizes a similarity structure analysis algorithm to extract structural patterns of PV output deviations, moving away from mean fitting or curve regression and establishing a structured identification path for deviation patterns. This allows for the output of differentiated compensation reference values ​​for different meteorological conditions, shifting energy storage control from "fluctuation tracking" to "deviation feedforward compensation."

[0061] Meanwhile, this invention extends the traditional energy storage cost model into a multi-objective optimization model with "annual comprehensive cost + output deviation dual objectives," and for the first time incorporates opportunity compensation cost into the optimization index system, making the optimization objectives more consistent with the economic constraints of actual operation. At the solution level, this invention constructs an improved multi-objective Harris Eagle optimization algorithm by introducing mechanisms such as dynamic learning factors, non-dominated sorting, and dynamic maintenance of Pareto boundaries. This achieves efficient search for non-stationary deviation characteristics and the ability to maintain the evolution of the optimal solution set. The optimization results not only achieve lower levelization costs but also simultaneously reduce the actual output deviation rate.

[0062] Furthermore, by constructing an adaptive implementation and online correction module during operation, this invention enables the energy storage scheduling model to automatically adjust parameters and compensation strategies in response to sudden changes in meteorological conditions. This achieves a closed-loop evolution of energy storage from "configuration to execution to monitoring to re-optimization," improving the system's robustness and applicability to environments with multiple meteorological disturbances. In summary, this invention addresses the problems of static energy storage configuration, delayed deviation compensation, and one-sided cost optimization in existing technologies, demonstrating significant technological advancement and engineering application value. Attached Figure Description

[0063] Figure 1 System framework diagram; Figure 2 Configuration scheme diagram; Figure 3 Energy storage charging and discharging diagram. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: like Figure 1-3 As shown, an optimal configuration method for energy storage systems based on real-time adaptive compensation for output deviation is proposed. This method dynamically extracts photovoltaic power deviation characteristics by introducing a similarity structure analysis algorithm and an opportunity compensation cost mechanism, establishes a deviation target domain, and solves the problem using the multi-objective Harris Hawk (HHO) algorithm to obtain the Pareto optimal configuration scheme for the capacity and number of energy storage systems, thereby achieving adaptive optimization of the energy storage system and minimization of output deviation.

[0065] Furthermore, the above method consists of the following steps: Step 1: Acquisition and Preprocessing of High-Altitude Solar Power Output Data (4) Data acquisition In response to the challenges of complex climate, difficult meteorological data sampling, and significant impact of terrain-induced shading on sunlight in high-altitude regions, this invention proposes a multi-source fusion method for acquiring high-altitude photovoltaic power output data, comprising the following: ① Real-time operation data collection of photovoltaic power plants The SCADA system collects real-time information on photovoltaic module string voltage, current, power, temperature, and combiner box status; the sampling period is set to 5 minutes and stored in time series format for short-term dynamic output analysis; the inverter output active power and bus voltage and current data are collected simultaneously to form a high-resolution photovoltaic output record.

[0066] ②Simultaneous acquisition of meteorological and geographic data Meteorological parameters such as solar irradiance, wind speed, wind direction, temperature, humidity, air pressure, snow thickness, and cloud cover are obtained through on-site meteorological stations. Digital elevation models (DEM) and terrain shading correction algorithms are used to calculate correction parameters for the angle and intensity of terrain shading.

[0067] ③ Integration of historical output and forecast data The system integrates long-term sunshine forecast data from meteorological departments with historical processed data; it uses time-synchronous interpolation to complete missing historical samples; and it constructs a unified photovoltaic power-meteorological parameter database to provide data support for subsequent deviation analysis and modeling.

[0068] (5) Data preprocessing To ensure the stability and analyzability of photovoltaic power output data, this invention employs a multi-level data cleaning and feature decomposition mechanism, with the following steps: ①Outlier detection and cleaning Outliers in the photovoltaic power sequence are identified using the three-σ criterion and box plot detection method; abnormal segments caused by sensor failure, inverter shutdown, and snow shading are marked; short-term outliers are corrected using the sliding window mid-range filtering method; and continuous missing measurement segments are reconstructed and completed using ARIMA.

[0069] ② Noise reduction and trend decomposition Wavelet packet decomposition (WPD) was used to perform multi-scale decomposition of the photovoltaic power series, separating the long-term trend term from the high-frequency noise term: the high-frequency component corresponds to random disturbances and measurement noise; the low-frequency component corresponds to climate change trends and solar cycle characteristics. Only the low-frequency subband signal was retained for long-term power output trend analysis to improve the model's stability.

[0070] ③ Normalization and Reconfiguration of Photovoltaic Output Characteristics Each characteristic parameter (irradiance, temperature, wind speed, etc.) is subjected to minimum-maximum normalization, with a range of 0 to 1; and a terrain correction factor is applied. The weighted correction to the normalization result is calculated using the following formula: , in, It is calculated from topographic factors such as slope, aspect, and shading angle.

[0071] ⑤ Dataset partitioning and feature annotation Based on the principle of temporal continuity, the data was divided into a training set (70%), a validation set (15%), and a test set (15%). Each sample data was labeled with its timestamp, corresponding meteorological state label (sunny, cloudy, snowy, foggy, etc.), and actual power deviation, providing an input basis for subsequent adaptive deviation analysis.

[0072] (6) Data feature extraction and modeling interface After preprocessing, feature extraction is performed on the cleaned photovoltaic output data, mainly including: Intraday fluctuation range of photovoltaic power output Power change rate Smoothness index Coupling coefficient between light and temperature Output periodicity index .

[0073] The extracted features are passed as input variables to the next step, the "Photovoltaic Output Deviation Adaptive Extraction Algorithm" module, for training the similarity structure analysis model and constructing the deviation target domain.

[0074] Step 2: Adaptive Extraction of Photovoltaic Output Deviation (4) Overall Approach Photovoltaic power output in high-altitude areas is affected by complex climate and terrain conditions, and the deviation between predicted and actual power output exhibits nonlinear, non-stationary, and multi-scale coupled characteristics. Traditional deviation modeling methods, such as linear regression and simple time moving average, cannot effectively capture the dynamic structural features of power output deviation. To address this, this invention proposes an adaptive extraction algorithm for photovoltaic power output deviation based on similarity structure analysis (SSA). This algorithm constructs a multidimensional similarity matrix between predicted and historical power output, dynamically identifies the statistical structural features of power deviation, and updates the deviation target domain in real time through an adaptive threshold mechanism, achieving accurate extraction and self-learning correction of photovoltaic power output deviation.

[0075] (5) Algorithm principle Assume the predicted photovoltaic output value at time t is The actual observed output is Its power deviation is defined as: , Deviation sequence under complex climatic conditions It exhibits multiple temporal correlations. This algorithm performs adaptive extraction through the following three core processes: ① Similarity structure modeling: Analyze the multidimensional similarity between predicted output and historical output, and extract the deviation features of similar time periods.

[0076] ② Deviation pattern recognition: Through clustering and structure mapping, the deviation distribution patterns under different meteorological conditions are identified.

[0077] ③ Adaptive update of target domain: The deviation target domain is dynamically adjusted according to the real-time deviation trend to realize the self-learning feature of the model.

[0078] (6) Algorithm implementation process ⑥ Construct a multidimensional feature matrix Transform photovoltaic power output data into a multi-dimensional input vector: , in, Solar irradiance; :temperature; Wind speed; :humidity; : Incident angle correction factor; : Predicted output value.

[0079] At the same time, define the corresponding actual processing vector: , Constructing the feature set: , Used for subsequent similarity calculations.

[0080] ⑦ Construction of similarity structure matrix The Gaussian kernel similarity function is used to measure the similarity between any two time points. Processing structural similarity: , in, The kernel width parameter is adaptively set based on the data standard deviation.

[0081] To improve the temporal sensitivity of similarity, a time decay weight is introduced: , Construct a weighted similarity matrix: , This matrix reflects the dynamic structural correlation of photovoltaic power output under different meteorological and temporal conditions.

[0082] ⑧ Biased clustering and structure mapping Through the Perform spectral clustering to divide all samples into A cluster of deviation patterns: , Each cluster represents a typical output deviation characteristic structure (e.g., high irradiation-low temperature region, low irradiation-high humidity region, etc.).

[0083] Within each cluster, calculate the corresponding mean deviation eigenvector: , Further establish the structural mapping relationship between "meteorological conditions and deviation models": , This mapping function is used to identify the power output deviation model under the current meteorological conditions in real time.

[0084] ⑨ Adaptive determination of the deviation target domain Define the target domain of deviation as: , To achieve real-time adaptive updates, this invention introduces a sliding time window mechanism (window length L) and dynamic threshold update rules: , in It is an adaptive learning rate.

[0085] This mechanism can quickly adjust the upper and lower limits of the deviation range under sudden meteorological changes (such as snow accumulation or cloud shadows), achieving dynamic self-learning.

[0086] ⑩ Deviation feature extraction and compensation signal output Finally, the real-time deviation compensation signal is calculated: , in, It is a real-time power compensation command for the energy storage system. It is a compensation coefficient that is dynamically adjusted based on historical performance.

[0087] The output results serve as input to the subsequent dynamic compensation model of the energy storage system, guiding the energy storage system to perform adaptive charge and discharge control.

[0088] Step 3: Establishment of Dynamic Compensation Model for Energy Storage System (4) Energy state modeling of energy storage system The dynamic characteristics of an energy storage system can be characterized by its state of energy (SOC). The rated capacity of the radio energy storage system is... E max， exist t The SOC at time t is defined as: , in E ( t The current energy level of the energy storage unit is denoted as SOC. The change in SOC over time is determined by the power input and output. , in, and These represent the discharge and charging power, respectively. These represent the discharge and charge efficiencies, respectively. This model allows for real-time calculation of the energy storage system's energy changes, ensuring it remains within the permissible State of Charge (SOC) range. Internal operation.

[0089] (5) Dynamic power response model of energy storage system To describe the response characteristics of an energy storage system to fluctuations in photovoltaic power output, a power dynamic model needs to be established. The energy storage system output power... Driven by the target power deviation signal and affected by system response delay and dynamic constraints: , in, This refers to the photovoltaic power output deviation obtained through the deviation extraction algorithm; This is the system gain coefficient, used to control the compensation strength; The time constant representing the power response of the energy storage system reflects the system's inertial characteristics. By discretizing this equation, dynamic simulation and control strategy solutions based on real-time data can be achieved.

[0090] (6) Design of dynamic compensation strategy Based on the deviation detection and power response model, a dynamic compensation strategy for the energy storage system is designed. The core idea is to adjust the energy storage power output in real time according to the amplitude and trend of the photovoltaic power output deviation, thereby achieving energy balance and output smoothing. Its control strategies can be divided into the following three categories: ⑤ Fast power compensation tracking: When photovoltaic power output fluctuates rapidly (such as sudden changes in cloud shadows), the energy storage system uses the power error as input and employs proportional-integral control: , This strategy can quickly suppress short-term power fluctuations and maintain output continuity.

[0091] ⑥ Slow time domain energy balance compensation: For continuous irradiance variations or diurnal power imbalances, energy allocation is based on the SOC regulation mechanism: , By gradually adjusting the charging and discharging power, the energy storage system can be maintained to operate within a safe energy range.

[0092] ⑦ Adaptive compensation optimization: Introducing a fuzzy logic mechanism to dynamically adjust based on historical deviation distribution and system operating status. These parameters enable adaptive energy regulation across multiple time scales.

[0093] ⑧ Model Implementation and Simulation Verification The established energy storage dynamic compensation model is embedded into the power balance equation of the high-altitude photovoltaic system, resulting in: , The model was simulated and validated using the MATLAB platform, which evaluated the impact of different compensation parameters and control strategies on the system output stability. The results show that the model can effectively reduce the fluctuation amplitude of high-altitude photovoltaic output and improve the system power smoothing coefficient and grid connection quality.

[0094] Step 4: Construction of Multi-Objective Optimization Model (4) Overall Model Concept The power output of high-altitude photovoltaic systems is significantly affected by complex climatic environments, exhibiting substantial fluctuations and uncertainties. Dynamic compensation through energy storage systems can improve power smoothness, but energy storage configurations under different capacities and control strategies can significantly impact system economics. Therefore, this model aims to find the optimal balance between economics and stability, achieving the joint determination of optimal energy storage capacity, power ratio, and operating strategy through a multi-objective optimization method.

[0095] (5) Optimize the design of the objective function Objective 1: Minimize the annual overall cost of the energy storage system Annual comprehensive cost of energy storage system Including investment depreciation costs, operation and maintenance costs, and energy loss costs over the lifespan, it can be expressed as: , in: Initial investment cost of energy storage equipment; Investment cost per unit capacity and power; Energy storage system lifespan (years); Operating and maintenance costs; Energy loss costs caused by energy storage round-trip efficiency; Round-trip power efficiency.

[0096] In addition, the levelized cost of electricity (LCOE) for energy storage is introduced as a long-term economic indicator: , in This represents the total discharge capacity of the energy storage system over its lifespan.

[0097] One of the optimization goals is to make Minimum.

[0098] Objective 2: Minimize photovoltaic power output deviation rate The deviation rate of photovoltaic output measures the system's smoothing ability and grid friendliness, and is defined as: , in: Combined output power of photovoltaic and energy storage systems; : Expected output power curve; N : Number of time steps.

[0099] To reflect the effectiveness of the system's operation, the following auxiliary performance indicators are also introduced: Energy storage system idle rate: , Reflects the utilization rate of energy storage devices; Energy storage load tracking error: , The final objective function in its synthesized form is: , in These are weighting coefficients used to balance various performance indicators.

[0100] (6) Constraints The optimization model must meet the physical and technical constraints of energy storage and system operation: ① Power constraint: , ② Energy constraint (SOC): , And satisfy energy balance: , ③System power balance: , ④ Service life constraints: The number of energy storage cycles shall not exceed the maximum allowable value: , Step 5: Solve using the multi-objective Harris Eagle optimization algorithm (1) Overall framework of the algorithm The Harris Eagle Optimization Algorithm (HHO) is a biomimetic swarm intelligence optimization algorithm that simulates the hunting behavior of falcon flocks for global search and local exploitation. To extend HHO into a multi-objective solver, this invention introduces the following improvements to the standard HHO framework: ① Dynamic Learning Factor (DLF) — Enables individuals to adaptively adjust step size and exploration / utilization balance during the iteration process to cope with the non-stationarity of photovoltaic power output deviation.

[0101] ② Non-dominated ranking strategy—used to rank candidate solutions hierarchically according to the merits of multiple objectives, and to maintain the diversity of solutions by combining crowding or distance indicators.

[0102] ③ Pareto boundary dynamic update mechanism—Introducing external archives to save the currently found Pareto solutions and dynamically cleaning and updating them to ensure the maintenance and evolution of high-quality solution sets.

[0103] (2) Decoding and feasible region Decision variable vector x: Represents variables related to energy storage configuration and operation, such as the rated energy of each energy storage unit. E i Rated power P i Number of units n i Lifetime options, backup capacity distribution, etc. If the system consists of... M If the energy storage is composed of various groups, then: , Variable range: Define upper and lower bounds for each variable. , (3) Objective function and constraint evaluation Given a decision vector x, compute the bi-objective. Combining constraints, an optimized fitness vector (for non-dominated ranking) is formed through a penalty mechanism: , Where g(x) represents the merging quantity of all inequality constraint violations. m A relatively large penalty coefficient (dynamically adjustable).

[0104] (4) Detailed design of key improvement points ① Dynamic Learning Factor (DLF) DLF is used to control the search stride and adaptive learning rate of each generation of falcons (individuals), and is defined as: , in t For the current iteration, T max The maximum number of iterations, k,g To adjust the parameters, DLF is kept at a large value in the early stages to enhance exploration, and then gradually reduced in the later stages to improve development accuracy; the noise term helps to cross local optima.

[0105] ② DLF integration in HHO action strategy In HHO's different hunting strategies (encirclement, ambush, encirclement, rapid pursuit, etc.), the step size in the individual update formula is multiplied by... ,For example: , ③ Non-dominated sorting + crowding (or distance) maintenance Perform a non-dominated sort on the solution set after merging the current population with external archives, and then stratify the solutions into fronts. F 1, F 2,…).

[0106] Within the same frontier, solutions are ranked using crowding distance or a homogeneity metric based on the target space to select candidates that maintain diversity for the next generation or archive.

[0107] Selection strategy: Prioritize retaining the more advanced non-dominated layers, and within each layer, select from the most crowded to the least crowded, until the population / archive capacity is full.

[0108] ④ Pareto boundary dynamic update mechanism Maintain a finite-capacity external archive A to store currently discovered non-dominated solutions. Each time a new solution is generated: merge the new solution with the solutions in A; delete dominated solutions; if the archive size exceeds the limit, remove redundant solutions based on crowding or target space uniformity to ensure Pareto boundary coverage and diversity; periodically perform a mutation to introduce slight variations into the archive to expand boundary exploration.

[0109] Archive updates use a dynamic threshold: when the improvement in the objective function value is less than a threshold... ϵ At the same time, the standards for archival retention should be tightened to improve the quality of data collection.

[0110] ⑤ Constraint handling strategy Main strategy: Repair + Penalty. For easily repairable constraints (such as integerization and power upper bounds), a repair function is used to directly map them to the feasible region; for difficult-to-repair or complexly coupled constraints, an adaptive penalty is employed. , The penalty is adaptively adjusted according to the proportion of constraints not met, prompting the search to gradually converge toward the feasible region.

[0111] Step 6: Adaptive Implementation and Online Correction (1) Real-time monitoring and data acquisition module The system deploys multi-source data acquisition terminals to monitor in real time the photovoltaic array output, grid-side power, energy storage SOC status, meteorological parameters (irradiance, temperature, wind speed, etc.) and load-side power demand of the high-altitude photovoltaic power station.

[0112] The sampled data undergoes preliminary noise reduction and outlier detection at edge computing nodes. Outliers are removed using moving average filtering and the triple σ criterion, ensuring the timeliness and reliability of the input data. The data acquisition cycle is adaptively adjusted according to the operating scenario, typically set to 1-5 minutes to balance monitoring accuracy and communication bandwidth.

[0113] (2) Dynamic update mechanism for the deviation target domain Using the photovoltaic power output deviation adaptive extraction algorithm established in step 2, the system continuously calculates the deviation distribution sequence between the predicted photovoltaic power output and the measured power output during operation.

[0114] Based on a sliding time window mechanism (e.g., the past 24 hours), similarity structure analysis (SSA) is used to re-identify the photovoltaic power output mode. The latest deviation statistical characteristics are input into the deviation target domain to update the model, and the power deviation boundary conditions are dynamically corrected. When the deviation target domain changes significantly, the system automatically triggers the adjustment of the energy storage scheduling strategy.

[0115] (3) Adaptive scheduling strategy for energy storage charging and discharging Based on the updated deviation target domain and real-time load demand, the system employs an adaptive scheduling strategy to dynamically control the charging and discharging power of the energy storage system. Specifically, this includes: Charging strategy: When the photovoltaic output is higher than the predicted value and the SOC is lower than the safe limit, the system prioritizes fast charging to absorb the excess energy; Discharge strategy: When the photovoltaic output is lower than the predicted value or the net load of the grid is higher than the target value, the system dynamically releases energy according to the deviation target domain; Smooth control: Introducing proportional-integral-derivative (PID) control combined with fuzzy logic to suppress frequent start-stop of energy storage and extend equipment life.

[0116] The system adopts an economic constraint based on opportunity compensation cost to conduct real-time assessment of energy loss and depreciation caused by frequent energy storage operations, thereby achieving dual optimization of economy and stability.

[0117] (4) Model parameter self-learning correction mechanism To enhance long-term adaptive capabilities, the system incorporates a lightweight self-learning module to periodically correct key model parameters (such as bias mapping coefficients, SOC dynamic thresholds, and opportunity cost weights) online. The recursive least squares (RLS) method and an improved particle filter (PF) algorithm are used to jointly update parameters, ensuring that the model maintains high-precision prediction and control performance even under abrupt changes in climate conditions (such as snow accumulation, strong winds, or low-temperature radiation attenuation).

[0118] (5) Closed-loop optimization and execution feedback After each operating cycle, the system feeds back the results, such as the energy storage operating status, photovoltaic deviation compensation effect, and economic indicators (e.g., levelized cost of equity, LCOE), to the optimization model module, forming a data closed loop. Through the periodically triggered MO-HHO algorithm re-optimization process, the system can quickly adjust the Pareto solution set without re-initialization, realizing the transformation from "static optimization" to "dynamic self-evolutionary optimization".

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions conceived without inventive effort should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. An energy storage system optimization configuration method based on real-time adaptive compensation of output deviation, characterized in that, This method introduces a similarity structure analysis algorithm and an opportunity compensation cost mechanism to dynamically extract photovoltaic power deviation characteristics, establish a deviation target domain, and solve the problem using the multi-objective Harris Hawk (HHO) algorithm to obtain the Pareto optimal configuration scheme for the capacity and number of energy storage systems, thereby achieving adaptive optimization of the energy storage system and minimization of output deviation.

2. The method of claim 1 wherein: This method consists of the following steps: Step 1: Acquisition and Preprocessing of High-Altitude Solar Power Output Data (1) Data acquisition In response to the challenges of complex climate, difficult meteorological data sampling, and significant impact of terrain-induced shading on sunlight in high-altitude regions, this invention proposes a multi-source fusion method for acquiring high-altitude photovoltaic power output data. (2) Data preprocessing To ensure the stability and analyzability of photovoltaic power output data, this invention employs a multi-level data cleaning and feature decomposition mechanism; (3) Data feature extraction and modeling interface After preprocessing, feature extraction is performed on the cleaned photovoltaic output data, mainly including: Intra-day fluctuation amplitude of photovoltaic power output Rate of power change Smoothness index Coupling coefficient of illumination and temperature Periodicity index of output ; The extracted features are passed as input variables to the next step, the "Photovoltaic Output Deviation Adaptive Extraction Algorithm" module, for training the similarity structure analysis model and constructing the deviation target domain; Step 2: Adaptive Extraction of Photovoltaic Output Deviation (1) Overall Approach Photovoltaic power output in high-altitude areas is affected by complex climate and terrain conditions, and the deviation between predicted and actual power output exhibits nonlinear, non-stationary, and multi-scale coupling characteristics. Traditional deviation modeling methods, such as linear regression and simple time moving average, cannot effectively capture the dynamic structural features of power output deviation. To address this, this invention proposes an adaptive extraction algorithm for photovoltaic power output deviation based on similarity structure analysis (SSA). This algorithm constructs a multidimensional similarity matrix between predicted and historical power output, dynamically identifies the statistical structural features of power deviation, and updates the deviation target domain in real time through an adaptive threshold mechanism, thereby achieving accurate extraction and self-learning correction of photovoltaic power output deviation. (2) Algorithm principle Let us assume that at time t the photovoltaic power output forecast value is , the actual observed output is , and the power deviation is defined as: , Under complex climate conditions, the bias sequence Present multiple time correlations; this algorithm is adaptively extracted through the following three core processes: ① Similarity structure modeling: Analyze the multidimensional similarity between predicted output and historical output, and extract the deviation features of similar time periods; ② Deviation pattern recognition: Through clustering and structure mapping, the deviation distribution patterns under different meteorological conditions are identified; ③ Adaptive update of target domain: The deviation target domain is dynamically adjusted according to the real-time deviation trend to realize the self-learning feature of the model; (3) Algorithm implementation process ① Construct a multidimensional feature matrix Transform photovoltaic power output data into a multi-dimensional input vector: ， wherein, : solar irradiance; : temperature; : wind speed; : humidity; : incidence angle modifier; : predicted power value; At the same time, define the corresponding actual processing vector: ， Constructing the feature set: ， Used for subsequent similarity calculation; ② Construction of similarity structure matrix The Gaussian kernel similarity function is used to measure the similarity between any two time points. Processing structural similarity: ， wherein is a kernel width parameter, set adaptively according to the standard deviation of the data; To improve the temporal sensitivity of similarity, a time decay weight is introduced: ， Construct a weighted similarity matrix: ， This matrix reflects the dynamic structural correlation of photovoltaic power output under different meteorological and temporal conditions; ③ Biased clustering and structure mapping By performing spectral clustering on the all samples are divided into clusters of bias patterns: ， Each cluster represents a typical output deviation characteristic structure (e.g., high irradiation-low temperature region, low irradiation-high humidity region, etc.). Within each cluster, calculate the corresponding mean deviation eigenvector: ， Further establish the structural mapping relationship between "meteorological conditions and deviation models": ， This mapping function is used to identify the power output deviation model under the current meteorological conditions in real time; ④ Adaptive determination of the deviation target domain The deviation target domain is defined as: ， To achieve real-time adaptive update, the application introduces a sliding time window mechanism (window length L) and a dynamic threshold update rule: ， wherein is the adaptive learning rate; This mechanism can quickly adjust the upper and lower limits of the deviation domain under the condition of meteorological mutation (such as snow and cloud shadow), and realize dynamic self-learning; ⑤ Deviation feature extraction and compensation signal output The final calculation of the real-time deviation compensation signal is: ， wherein, is a real-time power compensation instruction of the energy storage system, is a compensation coefficient, dynamically adjusted according to historical performance; The output result is used as the input of the subsequent dynamic compensation model of the energy storage system to guide the adaptive charging and discharging control of the energy storage system; Step 3: Dynamic compensation model of energy storage system (1) Energy state modeling of energy storage system The dynamic behavior of an energy storage system can be characterized by its state of charge (SOC); the rated capacity of an energy storage system is E max， At t the SOC at time t is defined as: ， wherein E ( t ) is the actual energy of the current energy storage unit; the change of SOC over time is determined by the power input and output: ， wherein, with Pdisch and Pcharge represent the discharging and charging power, respectively, Edisch and Echarge are the discharging and charging efficiency, respectively; through this model, the energy variation of the energy storage system can be calculated in real time, ensuring that it operates within the allowed SOC range . (2) Power dynamic response model of energy storage system To describe the response characteristics of energy storage system to photovoltaic output fluctuation, it is necessary to establish a power dynamic model; the output power of energy storage system is driven by the target power deviation signal and is affected by system response delay and dynamic constraints: ， wherein, is the photovoltaic power output deviation amount obtained by the deviation extraction algorithm; is a system gain coefficient for controlling the compensation intensity; represents the time constant of the power response of the energy storage system, reflecting the inertia characteristics of the system; through the discretization processing of the equation, dynamic simulation and control strategy solving based on real-time data can be realized; (3) Dynamic compensation strategy design On the basis of deviation detection and power response model, the dynamic compensation strategy of the energy storage system is designed; the core idea is to adjust the power output of the energy storage system in real time according to the amplitude and trend of the photovoltaic output deviation, so as to realize energy balance and output smoothing; Step 4: Construction of multi-objective optimization model (1) Overall idea of model The output of high mountain photovoltaic system is affected by complex climate environment, showing significant volatility and uncertainty; through the dynamic compensation of the energy storage system, the power smoothness can be improved, but the configuration of the energy storage system with different capacity and control strategy will significantly affect the economy of the system; therefore, the model aims to seek the optimal balance point between economy and stability, and realizes the joint determination of the optimal energy storage capacity, power ratio and operation strategy through the multi-objective optimization method; (2) Design of optimization objective function Objective 1: Minimization of annual comprehensive cost of energy storage system Annualized cost of energy storage system The annualized cost of energy storage system (ACOS) includes the capital cost, operation and maintenance cost, and energy loss cost over the lifetime of the system, which can be expressed as: ， wherein: : initial investment cost of energy storage device; : unit capacity and power investment cost; : Energy storage system lifetime (years); : operating and maintenance costs; : Energy loss cost due to round trip efficiency of energy storage : round trip efficiency; In addition, the energy storage levelization electricity cost (LCOE) is introduced as a long-term economic index: ， wherein Qtotal is the total amount of energy discharged by the energy storage system over its lifetime; One of the optimization goals is to make the minimum; Objective 2: Minimization of photovoltaic output deviation rate The deviation rate of photovoltaic output measures the smoothness of the system and the friendliness to the power grid, and is defined as: ， Wherein: : combined output power of the light storage : desired output power curve; N : time steps; In order to reflect the operation effectiveness of the system, the following auxiliary performance indicators are introduced: Idle rate of energy storage system: ， Reflects the utilization degree of energy storage device; Energy storage load tracking error: ， The final target function is in the form of: ， wherein are weight coefficients for balancing the performance indicators; (3) Constraint conditions The optimization model needs to meet the physical and technical constraints of energy storage and system operation: ① Power constraint: ， ② Energy constraint (SOC): ， And meet the energy balance: ， ③ System power balance: ， ④ Operation life constraint: The number of cycles of the energy storage does not exceed the maximum allowed value: ， Step 5: Solution of multi-objective Harris hawk optimization algorithm (1) Overall framework of algorithm Harris hawk optimization algorithm (HHO) is a kind of bionic swarm intelligence optimization algorithm, which simulates the hunting behavior of hawk group for global search and local development; in order to extend HHO into a multi-objective solver, the following improvements are introduced on the basis of standard HHO framework: ① Dynamic learning factor (DLF) - makes individuals adaptively adjust step size and exploration / exploitation balance in the iteration process to cope with the non-stationarity of photovoltaic output deviation; ② Non-dominated relationship sorting strategy - used to sort the candidate solutions according to the multi-objective optimization, and combined with the congestion or distance index to maintain the diversity of solutions; ③ Dynamic updating mechanism of Pareto boundary - introduces an external archive to save the Pareto solutions found currently and dynamically clean and update, to ensure the maintenance and evolution of high-quality solution set; (2) Solution coding and feasible region Decision variable vector x: represents storage configuration and operation related variables, e.g. rated energy of each storage unit E i , rated power P i , number of units n i , lifetime options, reserve capacity distribution, etc.; if the system consists of M storage groupings, then: ， Variable range: define upper and lower bounds for each variable, , (3) Objective function and constraint evaluation Given the decision vector x, the double objective is calculated; combined with the constraints, the optimized fitness vector is formed through the punishment mechanism (for non-dominated sorting): ， where g(x) represents the aggregate amount of all inequality constraint violations, μ is a large penalty coefficient (dynamically adjustable); (4) Improvement details ① Dynamic learning factor (DLF) DLF is used to control the search step size and adaptive learning rate of each generation of falcon (individual), defined as: ， wherein t is the current iteration, T max is the maximum number of iterations, κ,γ is a tuning parameter; DLF keeps a larger value in the early stage to enhance exploration, and gradually decreases in the later stage to improve development accuracy; the noise term helps to cross the local optimum; ② Integration of DLF in HHO action strategy In different hunting strategies (encirclement, raid, surround, rapid pursuit, etc.) of HHO, the step in the individual update formula is multiplied by , for example: ， ③ Non-dominated sorting + crowding degree (or distance) maintenance The solution set after merging the current population with the external archive is non-dominantly sorted, and the solutions are layered (Fronts F 1, F 2, …); Within the same front, the solutions are sorted by crowding distance or uniformity measure based on the target space to select candidate individuals that maintain diversity into the next generation or archive; Selection strategy: prefer to keep the non-dominated layer closer to the front, and select by crowding degree from large to small within the layer until the population / archive capacity is full; ④ Dynamic updating mechanism of Pareto boundary Maintain a limited capacity external archive A to save the current discovered non-dominated solution; each time a new solution is generated: combine the new solution with the solution in A; delete the dominated solution; if the archive size exceeds the upper limit, remove redundant solutions according to the crowding degree or target space uniformity to ensure the coverage and diversity of the Pareto boundary; periodic disturbance (mutation) can be performed to introduce slight variation in the archive to expand the boundary exploration; Archive update with dynamic threshold: when the improvement of the objective function value is less than a threshold ϵ the archive retention criteria is tightened to improve the solution set quality; ⑤ Constraint processing strategy Main strategy: repair + penalty; for easily repaired constraints (such as integerization, power upper limit), use repair function to directly map to feasible domain; for difficult to repair or complex coupled constraints, use adaptive penalty: ， Make the penalty adaptively adjust with the proportion of not meeting the constraints, and promote the search to gradually converge to the feasible domain; Step 6: Adaptive implementation and online correction (1) Real-time monitoring and data acquisition module The system deploys multiple source data acquisition terminals to monitor the photovoltaic array output of high mountain photovoltaic power station, grid side power, energy storage SOC state, meteorological parameters (irradiance, temperature, wind speed, etc.) and load side power demand in real time; The sampled data is preliminarily denoised and outlier detected by the edge computing node, and the outliers are removed by using the moving average filter and three sigma criterion to ensure the timeliness and reliability of the input data; the data acquisition period is adaptively adjusted according to the running scene, generally set to 1~5 minutes to balance the monitoring accuracy and communication bandwidth; (2) Dynamic updating mechanism of deviation target domain Using the photovoltaic output deviation adaptive extraction algorithm established in step 2, the system continuously calculates the deviation distribution sequence between the photovoltaic prediction output and the measured output during operation; Based on the sliding time window mechanism (such as the past 24 hours), the similarity structure analysis (SSA) is used to re-identify the photovoltaic output mode, and the latest deviation statistical characteristics are input into the deviation target domain updating model to dynamically correct the power deviation boundary conditions; when the deviation target domain changes significantly, the system automatically triggers the energy storage dispatching strategy adjustment; (3) Adaptive dispatching strategy of energy storage charging and discharging According to the updated deviation target domain and real-time load demand, the system uses an adaptive dispatching strategy to dynamically control the charging and discharging power of the energy storage system; specifically including: Charging strategy: when the photovoltaic output is higher than the predicted value and the SOC is lower than the safe upper limit, the system preferentially executes fast charging to absorb excess power; Discharge strategy: when the photovoltaic output is lower than the predicted value or the grid net load is higher, the system releases energy dynamically according to the deviation target domain; Smooth control: introduce proportional-integral-derivative (PID) combined with fuzzy logic control to suppress the frequent start-stop of energy storage and prolong the service life of the equipment; The system uses economic constraints based on opportunity compensation cost to evaluate the energy loss and depreciation caused by frequent operation of energy storage in real time, achieving dual optimization of economy and stability; (4) Model parameter self-learning correction mechanism To improve the long-term operation self-adaptive ability, the system embeds a lightweight self-learning module to periodically correct the key parameters of the model (such as deviation mapping coefficient, SOC dynamic threshold, opportunity cost weight, etc.). Recursive least squares (RLS) and improved particle filter (PF) are used to update the parameters to ensure that the model can maintain high-precision prediction and control performance under sudden changes in climate conditions (such as snow, strong wind, or low-temperature radiation attenuation); (5) Closed-loop optimization and execution feedback After each operation cycle of the system, the energy storage operation state, photovoltaic deviation compensation effect, and economic indicators (such as levelized cost of electricity (LCOE)) are returned to the optimization model module to form a data closed loop. Through the periodic triggering of the MO-HHO algorithm re-optimization process, the system can quickly adjust the Pareto solution set without re-initialization, realizing the transition from "static optimization" to "dynamic self-evolution optimization".

3. The method of claim 2, wherein, The step 1: The multi-source fusion high-altitude photovoltaic output data acquisition method includes the following contents: ① Real-time operation data acquisition of photovoltaic power station Through the SCADA system, real-time acquisition of photovoltaic module string voltage, current, power, temperature, and bus status information is performed. The sampling period is set to 5 minutes, and the data is stored in time series format for short-term dynamic output analysis. Synchronous acquisition of inverter output active power and bus voltage, current data is performed to form a high-resolution photovoltaic output record; ② Synchronous acquisition of meteorological and geographical data Through the on-site meteorological station, meteorological parameters such as solar radiation, wind speed, wind direction, temperature, humidity, air pressure, snow depth, and cloud cover are acquired. Using a digital elevation model (DEM) and terrain shading correction algorithm, the correction parameters for the angle and intensity of the terrain are calculated; ③ Integration of historical output and prediction data Long-term sunshine prediction data and historical processing data from the meteorological department are called for integration. Time-synchronous interpolation method is used to complete the missing samples. A unified photovoltaic power-meteorological parameter database is constructed to provide data support for subsequent deviation analysis and modeling.

4. The method of claim 3, wherein, In the data preprocessing of step 1 (2), the multi-level data cleaning and feature decomposition mechanism is as follows: ① Outlier detection and cleaning Three-sigma criterion and box plot detection method are used to identify outliers in photovoltaic power sequence. Abnormal sections caused by sensor failure, inverter shutdown, and snow shading are marked. Short-term outliers are corrected using sliding window median filtering method, and continuous missing sections are reconstructed and completed using ARIMA; ② Denoising and trend decomposition The photovoltaic power sequence is decomposed into multiple scales using wavelet packet decomposition (WPD), separating the long-term trend from the high-frequency noise: the high-frequency part corresponds to random disturbances and measurement noise; the low-frequency part corresponds to climate change trends and sunshine cycle characteristics; Only the low-frequency sub-band signal is retained for long-term output trend analysis to improve the stability of the model; ③ Photovoltaic output feature normalization and reconstruction The minimum-maximum normalization is performed on each characteristic parameter (irradiance, temperature, wind speed, etc.), and the range is set to 0-1; and the terrain correction coefficient is obtained according to the following formula The normalized result is weighted and corrected, and the calculation formula is as follows: ， wherein, are calculated from terrain factors such as slope, aspect, and shading angle. ④ Dataset division and feature labeling According to the principle of time continuity, the data is divided into training set (70%), validation set (15%) and test set (15%); each sample data is labeled with its timestamp, corresponding meteorological state label (sunny, cloudy, snow, fog, etc.) and actual power deviation, providing input basis for subsequent deviation adaptive analysis.

5. The method of claim 4, wherein, In the step 3 (3) dynamic compensation strategy design, the control strategy can be divided into the following three categories: ① Fast power compensation tracking: When the photovoltaic output appears rapid fluctuation (such as cloud shadow mutation), the energy storage system takes power error as input, and adopts proportional integral control: ， This strategy can quickly suppress short-term power mutation and maintain output continuity; ② Slow-time energy balance compensation: For continuous irradiance changes or day-night power imbalance, energy allocation is based on SOC adjustment mechanism: ， By gradually adjusting the charge and discharge power, the energy storage system is maintained within the safe energy range; ③ Adaptive compensation optimization: The fuzzy logic mechanism is introduced to dynamically adjust the historical deviation distribution and system operation state and parameters to realize adaptive energy regulation of multiple time scales. ④ Model implementation and simulation verification The built energy storage dynamic compensation model is embedded in the power balance equation of high mountain photovoltaic system, forming: ， Through MATLAB platform simulation verification, the influence of different compensation parameters and control strategies on system output stability can be evaluated; the results show that the model can effectively reduce the fluctuation amplitude of high mountain photovoltaic output, improve the system power smoothing coefficient and grid quality.

Images