Real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-target scheduling method

Abstract

The invention provides a real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-target scheduling method, and belongs to the technical field of reservoir optimal scheduling. According to the method, a collaborative mechanism of data layer cross-scale fusion, modeling layer dynamic mapping, algorithm layer dual-mode optimization, risk layer extreme management and control and landing layer closed-loop correction is constructed aiming at a reservoir multi-target scheduling contradiction in a new energy access scene, and a flood algorithm is improved to design dual-mode adaptive logic to adapt to different time scale requirements. Through combination of extreme scene risk management and control and closed-loop feedback correction, water-light complementary peak regulation response delay is reduced, the ecological flow time sequence compliance rate is improved, and the flood control risk rate and the power generation loss are reduced. The method effectively breaks through the bottlenecks of cross-scale coupling fracture, insufficient algorithm adaptation and risk quantification deficiency in the prior art, is high in engineering practicability, and can be directly connected with an existing reservoir dispatching system.

Description

Technical Field

[0001] This invention relates to the field of reservoir optimization scheduling, and in particular to a real-time, high-frequency, cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method. Background Technology

[0002] With the development of the social economy and the upgrading of water resource management needs, reservoir scheduling needs to take into account multiple objectives such as flood control, power generation, water supply, and ecology. In addition, the large-scale access of new energy sources such as photovoltaics has added the demand for water-solar complementary peak shaving, making the scheduling scenario more complex.

[0003] Existing reservoir multi-objective scheduling technologies suffer from problems such as cross-scale coupling breaks, insufficient algorithm time-scale adaptation, and inadequate risk quantification and extreme scenario adaptation. Existing technologies mostly rely on low-frequency data at the daily and ten-day levels, and have not established a correlation mechanism between hourly high-frequency data and multi-objective weights and risk quantification. This leads to subjective weight allocation and blind risk management in dynamic scenarios. For example, CN113935603B uses fixed flow level threshold coding, which cannot adapt to the high-frequency fluctuations in new energy output and the pulse demand of ecological flow. Traditional metaheuristic algorithms, such as the improved chimpanzee algorithm of CN119168263A, use fixed parameter optimization, which makes it difficult to balance the real-time response requirements of peak shaving with the global optimal accuracy of daily and ten-day level conventional scheduling, resulting in an inherent contradiction between convergence speed and solution set quality. Meanwhile, the existing technology lacks a closed-loop feedback mechanism, and after long-term rolling optimization, the deviation between the operating parameters and the actual hydrological scenario gradually amplifies, affecting the accuracy of project implementation.

[0004] Therefore, there is an urgent need for a multi-objective scheduling method for reservoirs that can achieve cross-scale collaboration, dynamic adaptive optimization, precise risk control and closed-loop correction, in order to solve the technical problems in the scenario of new energy access. Summary of the Invention

[0005] The main objective of this invention is to provide a real-time, high-frequency, cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method to solve the problems of adapting to sudden water inflows, new energy peak shaving and global optimal accuracy, and controllable ecological risks in reservoir multi-objective scheduling under new energy access scenarios. It also breaks through the technical bottlenecks of risk coupling breakage of cross-scale data and insufficient time scale adaptation of algorithms.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a real-time, high-frequency, cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method. The method includes the following steps: S1. Collect multi-dimensional and multi-frequency scheduling basic data, and integrate it into cross-scale input data of high-frequency dynamic features and low-frequency trend features through preprocessing and fusion operations; S2. Establish a coupled mapping mechanism for cross-scale input data mapping and multi-objective dynamic weight mapping for flood control / ecological risk quantification, with subsequent steps providing objective guidance and risk constraints; S3. Based on the state of the order parameter and the dynamic weight, identify the current scheduling scenario, improve the flood algorithm (FLA) and design a dual-mode adaptive logic to adapt to real-time peak shaving and conventional optimization scenarios respectively; S4 identifies coupled extreme scenarios such as heavy rain, sudden drop in new energy sources, ecological pulse, and insufficient inflow. It fills the risk coverage gap of the conventional dual-mode by adjusting risk redundancy and emergency optimization of FLA. S5. Transform the solution vector obtained by FLA optimization into an operable reservoir scheduling and control scheme to realize the engineering practicality of the scheme. S6. Construct a closed-loop mechanism to eliminate the cumulative deviation between operating parameters and actual hydrological scenarios in real time, ensure the continuous effectiveness of the scheduling scheme, and avoid the amplification of deviations caused by long-term rolling optimization.

[0007] In the preferred embodiment, in step S1: Data acquisition employs a distributed sensor fusion platform architecture, covering both high-frequency dynamic data and low-frequency basic data, including: High-frequency data includes inflow data, renewable energy output data, and ecological flow pulse demand data: Inflow data is collected by installing multiple ultrasonic flow sensors at the reservoir inlet to periodically collect instantaneous flow data. The average value of multiple sets of data is taken as the inflow value for that hour. At the same time, data such as flow fluctuation amplitude and rate of change are recorded. The power output data of new energy sources is collected periodically through the data acquisition and monitoring system of the corresponding power station. At the same time, environmental data is collected simultaneously for subsequent anomaly detection. Ecological flow pulse demand data is obtained by deploying an acoustic Doppler current profiler downstream of the reservoir outlet to collect outflow data periodically; combined with the watershed ecological protection plan, ecological flow pulse demand thresholds for different time periods are preset to form an ecological flow pulse demand dataset. Low-frequency data includes trend features and trend characteristics: The flood control limit reservoir capacity data is determined based on the reservoir design specifications and the annual flood control scheduling plan. The flood control limit water level for the year is determined and the flood control limit reservoir capacity is calculated in combination with the reservoir capacity curve, which serves as low-frequency constraint data. The long-term power generation target data is broken down into monthly and ten-day power generation targets based on the annual power supply plan of the power grid and the water inflow forecast of the reservoir. Other basic data include reservoir water level, reservoir capacity curve, water level and power output relationship curve, downstream flood control control water level, etc., as static basic data.

[0008] Because high-frequency data is susceptible to outliers due to sensor malfunctions and environmental interference, the 3σ criterion is used for outlier removal. First, the high-frequency data is grouped by type, with each group using a sliding window of several days. For the data within each sliding window, the mean μ and standard deviation σ are calculated. (1) (2) in, The number of data items within the sliding window; For the first One data point; For each data point within the window If satisfied If it is, then it is determined to be an outlier.

[0009] Furthermore, dynamic information is extracted from high-frequency data after outlier removal, while trend characteristics are extracted from low-frequency data, including: High-frequency data features: By calculating the mean, maximum, minimum, variance, skewness, and kurtosis within a sliding window for each type of high-frequency data, the central tendency and fluctuation characteristics of the data can be obtained. The dynamic trend of data change is obtained by calculating the rate of change, cumulative change, and impulse factor. By calculating the cross-correlation between inflow and new energy output, the coordinated fluctuation pattern between the two can be identified. In low-frequency data characteristics: By using linear fitting to extract the trend slope from the power generation target data during the flood season, the increasing or decreasing trend of the power generation target can be determined. By converting constraint data such as flood control reservoir capacity and downstream flood control water level into feature factors, it is easier to use in subsequent models. The high and low frequency features mentioned above are used to construct a feature extraction module using the Pandas and Scikit-learn libraries. The module automatically calculates the above features by inputting the standardized raw data, forming a two-layer dataset, expanding the feature dimensions, and providing information support for cross-scale fusion.

[0010] Furthermore, a hierarchical dynamic factor model is employed to fuse high-frequency feature data with low-frequency data. The model includes an input layer, a factor layer, and an output layer, wherein: The input layer is divided into a high-frequency input layer and a low-frequency input layer; The factor layer includes trend factor, volatility factor and synergy factor. The trend factor captures the long-term trend of low-frequency data, the volatility factor captures the short-term volatility of high-frequency data, and the synergy factor captures the correlation characteristics between high- and low-frequency data. The output layer outputs the fused cross-scale input data, which includes trend and fluctuation components.

[0011] The model selects historical data and upscales low-frequency data to high frequencies using interpolation to form an aligned training dataset. Then, it uses maximum likelihood estimation to estimate the model parameters and solves for the optimal parameter solution using the expectation-maximization algorithm until the update magnitude is less than the preset value. Finally, it verifies the model fusion effect using root mean square error and mean absolute percentage error. If the results are not satisfactory, the number of factors is adjusted and the model is retrained.

[0012] High- and low-frequency data are obtained by inputting high-frequency feature data of the current hour and up-frequency feature data into a trained hierarchical dynamic factor model, and outputting fused cross-scale input data: (3) (4) in, for The output data after time-fusion, where Λ is the factor loading matrix. for The dynamic factor vector at time step, Let Φ be the observation error, and Φ be the factor autoregression coefficient matrix. For factor error; This provides input data for subsequent cross-scale coupling mapping.

[0013] In the preferred embodiment, in step S2: In cross-scale sequence parameter identification, the cross-scale sequence parameters are the fluctuation range of inflow and the rate of change of new energy output, among which: Inbound flow fluctuation range This is the absolute value of the difference between the current hourly inbound flow and the average inbound flow over several hours, expressed by the formula: (5) in, Let t be the inbound flow rate. For the inbound flow at time ti, For comparison with the past; Rate of change in new energy output The formula for the relative change in renewable energy output between the current hour and the previous few hours is: (6) in, To contribute to new energy sources at time t; If it contributes to the new energy source at time t-1, Then take The absolute value is taken as the rate of change; Based on high-frequency data from the past few years, the order parameter thresholds are categorized into three types—low volatility, medium volatility, and high volatility—using the K-means clustering algorithm. The convergence condition is that the change in cluster centers is less than a preset value. The minimum threshold of the high volatility category is taken as the warning threshold. High volatility is defined as the fluctuation amplitude of inflow exceeding a preset value, and the change rate of new energy output exceeding a preset value. Simultaneously, the order parameters for each hour are also considered. , Compared with the warning threshold, the status is divided into normal, single-factor high volatility, and double-factor high volatility.

[0014] A cross-scale mapping model is constructed using the Bayesian dynamic factor algorithm. Model structure design: The input variable X includes high-frequency features and low-frequency features; There are four latent variables F, which are the weighted driving factors of the four objectives of flood control, power generation, ecology, and peak regulation. The output variable W is the weight vector W=[w1,w2,w3,w4] of the four objectives, which satisfies w1+w2+w3+w4=1, where w1 is the flood control weight, w2 is the power generation weight, w3 is the ecological weight, and w4 is the peak shaving weight.

[0015] The observation equation is: (7) in Let be the input vector at time t, and Λ be the factor loading matrix. Let be the latent variable vector at time t. This is the observation error; The state equation is: (8) Where Φ is the autoregressive coefficient matrix. This refers to the state error; The weighted output equation is: (9) Where A is the weight transformation matrix and b is the bias vector; The softmax function ensures that the sum of the weights is 1, and the formula is: (10) in, Let be the weight of the i-th target at time t; In model training, historical cross-scale input data X and corresponding manually labeled weights W are selected to construct the training dataset; First, set a normal prior N for the factor loading matrix Λ and the autoregressive coefficient matrix Φ, and then set the error variance... , Set the inverse Gamma prior (IG); The MCMC sampling solution uses the Metropolis-Hastings algorithm to perform MCMC sampling. A suitable number of samplings is discarded during the warm-up period to eliminate the influence of initial values. The remaining samplings are used for parameter estimation. This process is repeated until the potential scaling factor is less than the preset value. Finally, the confusion matrix and weight prediction error are used to evaluate the model performance: (11) in, Here, n represents the weight of the i-th target predicted by the model at time t, and n is the number of validation data points. when When the value is less than the preset value, the model performance meets the requirements.

[0016] Dynamic weight adjustment is based on the order parameter state and BDF model prediction results, incorporating cross-scale input data from the current hour. Input the trained BDF model to obtain the basic weights: ; when , Compared with the warning threshold, if the condition is determined to be normal, the basic weight remains unchanged, and the output is: ; When only During periods of high volatility, increase the flood control weight w1 and decrease the power generation weight w2. The adjustment formula is as follows: (13) (14) in for Fluctuation threshold; for Fluctuation threshold; When only During periods of high volatility, increase the peak-shaving weight w4 and decrease the power generation weight w2. The adjustment formula is as follows: (15) (16) when , When the average fluctuation is high, simultaneously increase the flood control weight w1 and the peak-shaving weight w4, and decrease the power generation weight w2. The adjustment formula is as follows: (17) (18) - (19) When an ecological pulse demand scenario occurs, the ecological weight w3 is increased, and the corresponding proportion is deducted from the power generation weight w2 to maintain the weight sum at 1. In the event of a flood control emergency, the flood control weight w1 will be increased to the maximum level, and the remaining weights will be allocated proportionally to maintain a total weight of 1.

[0017] Furthermore, based on cross-scale input data and dynamic weights, probabilistic statistical methods are used to quantify flood risk rate and ecological risk rate. Define flood risk rate To determine the probability that the reservoir discharge will exceed the downstream flood control safety discharge during the forecast period, a future time frame is selected as the forecast period, and the inflow during that time frame is predicted. The inflow rate is predicted using a long short-term memory network model. The model input consists of past inflow rate data and weather forecast data, and the output is the predicted inflow rate and 95% confidence interval for a set future time. The calculation method is based on the confidence interval of the flow forecast and uses the Monte Carlo simulation method to calculate the flood risk rate. The specific method is as follows: Randomly sample the inbound flow sequence from the normal distribution of flow forecast; For each sampling sequence, combined with the current storage capacity Flood control weight in dynamic weighting The sequence of discharge flow at a predetermined time in the future is calculated using the water balance equation. : (20) in, ; Let k be the storage capacity. Let k be the evaporation and seepage loss at time k; The outflow exceeded the downstream flood control safety discharge capacity. The number of sampling groups m is used to calculate the flood control risk rate; Ecological risk rate Defined as the probability that the outflow from the reservoir during the forecast period will not meet the ecological flow pulse demand; based on the ecological protection plan, the ecological flow demand range for a given time period in the future is determined. Using the Monte Carlo simulation method, which is the same as the flood control risk rate, the number of outflow groups n in the sampled array sequence is counted to calculate the ecological risk rate.

[0018] In the preferred embodiment, in step S3: The improved flood algorithm (FLA) is designed with dual-mode adaptive logic. Based on the order parameter state and dynamic weight, it identifies the current scheduling scenario, treats each solution vector as a water flow unit, and the dimension of the solution vector corresponds to the scheduling decision variable. A set of water flow units with a size of N is randomly generated. New solutions are generated through random perturbation, expanding the search space. The formula is as follows: ;(twenty one) in, This is the old interpretation. For the new solution, α is the diffusion coefficient. These are random numbers from a standard normal distribution. , The upper and lower boundaries of the solution; The formula for performing a local refinement search on the high-quality solution obtained from the global search is as follows: ;(twenty two) in, This represents the current optimal solution, where β is the permeability coefficient. The result is a uniformly random number in the range [0,1]. The non-dominated sorting and crowding degree calculation are used to screen high-quality solutions, retain non-dominated solutions, and eliminate inferior solutions. If the change in HV volume of the non-dominated solution set for N consecutive generations is less than the preset value, or if the maximum number of iterations is reached, the iteration stops and the optimal solution is output.

[0019] Scene recognition is based on order parameter states and peak-shaving weights. Scene determination metrics include: The core indicator is peak shaving weight. Change rate of power output with new energy sources ; The auxiliary indicator is the fluctuation range of the predicted output of new energy sources within a future set time period; When satisfied Peak shaving weight is higher than the preset value or A value exceeding the preset value is identified as a real-time peak shaving scenario. If the above real-time peak shaving scenario conditions are not met, it is judged as a regular optimization scenario; During the scene recognition process, scene recognition is performed at a fixed frequency. If the scene changes, the FLA algorithm parameters are automatically adjusted to adapt the algorithm to the new scene.

[0020] For different scenarios, the FLA parameters and execution logic are optimized differently. When in a real-time peak shaving scenario: The dimension of the solution vector X is defined by future time, with two decision variables set for each hour: water level adjustment amount and... The water discharge volume is ; Water level adjustment amount Set boundary values ​​according to the situation to avoid drastic adjustments; Water discharge The lower limit is set as the lower limit of ecological flow, and the upper limit is set as the maximum discharge of the reservoir; FLA settings include: population size N; maximum number of iterations; diffusion coefficient α; penetration coefficient β; and crossover probability. ; Adjust appropriate values ​​to weaken local search and focus on rapid convergence to accelerate the spread of high-quality genes; The fitness function, guided by dynamic weights and incorporating risk constraints, is F(X): ;(twenty three) in Adaptability to flood control objectives; Adaptability to power generation targets For ecological goals of fitness; The peak-shaving target adaptability reflects the complementarity between the fluctuations in renewable energy output and the fluctuations in reservoir power generation; γ is the risk penalty coefficient. The algorithm first initializes the population by randomly generating solution vectors within the variable boundaries to form the initial population. Then, for each solution vector, a new solution is generated according to the diffusion formula, and the fitness function value is calculated. Next, for the solution vectors with high fitness, a local search is performed according to the penetration formula to update the solution vectors. Offspring solutions are generated using single-point crossover and polynomial mutation, and the solution vectors with high fitness are retained. Finally, a convergence judgment is performed, the iteration stops, and the solution vector with the highest fitness is output as the optimization result. In a normal scheduling scenario: The dimension of the solution vector X is defined as a longer future setting time, with two decision variables per hour consistent with the implementation of peak shaving scenarios; The variable boundary constraints are consistent with the implementation of peak shaving scenarios; FLA parameter settings are consistent with the implementation of peak shaving scenarios, with specific parameters biased towards globality to ensure search diversity and balance global search with local optimization: The fitness function design is also consistent with the implementation of peak shaving scenarios; By introducing dynamic crossover mutation probability adjustment, the HV hypervolume of the current generation of non-dominated solution set is calculated; If HV is higher than the preset value, the parameters remain unchanged to maintain the stability of the solution set; If HV is less than the preset value, the mutation probability is increased, the crossover mutation intensity is increased, and the search range is expanded. The algorithm execution steps are consistent with those for implementing peak shaving scenarios, but the number of iterations is increased and the convergence judgment conditions are tightened to enhance accuracy.

[0021] The performance optimization of FLA was dynamically evaluated using HV hypervolume, GD convergence distance, and SP distribution. The HV hypervolume is used to evaluate the convergence and distribution of non-dominated solution sets. A larger HV value is better. The formula is: ;(twenty four) Where P is the set of non-dominated solutions. Let x be the objective function vector, and λ be the Lebesgue measure; The GD convergence distance is used to evaluate the distance between the non-dominated solution set and the true Pareto front. A smaller GD value is better. The formula is: (25) in, For the true forefront of Pareto, To solve for the m-th objective function value of x, The total number of objective function values; The SP distribution property is used to evaluate the uniformity of the distribution of non-dominated solution sets. A smaller SP value is better. The formula is: (26) in, The minimum Euclidean distance between solution x and all other solutions. For all The average value; At the same time, different indicator thresholds are assigned to different scenarios as iteration termination conditions.

[0022] In the preferred embodiment, in step S4: A fusion model that combines ordered parameter states with risk rate thresholds for logical judgment is used to identify coupled extreme scenarios. During the recognition process: Step 1: Obtain the state of the order parameter output in Step 2 , With risk rate , ; Step 2: Obtain the current ecosystem traffic demand And the level of inbound flow; Step 3: Perform logical judgment according to the definition of extreme scenarios. If the preset extreme conditions are met, it is determined to be a coupled extreme scenario.

[0023] To address the uncertainty in hydrological forecasting under extreme scenarios, a Box fusion polyhedral set is used for quantization, where: The Box collection is defined as: (27) in This represents the maximum range of prediction error. This is the hydrological forecast value; Assuming a 95% confidence interval, a polyhedral set is defined as follows: (28) in The critical value for a chi-square distribution with 1 degree of freedom and a 95% confidence level; Uncertainty propagation combines the Box set with the polyhedral set to obtain the uncertain set of inbound traffic: (29) Based on the water balance equation, the discharge flow rate for: (30) in For the current storage capacity, For the target storage capacity, This refers to losses due to evaporation and leakage. Therefore, the formula for calculating the maximum possible value of the flood risk rate under extreme scenarios, \(R_{f,max}\), can be derived as follows: (31) in This is the minimum safe storage capacity of the reservoir; To adjust the flood control limit reservoir capacity; The output is a probability function, indicating that the outflow exceeds the flood control safety outflow. The probability of; Ensure that the propagation of uncertainty conforms to the actual operating boundaries of the reservoir.

[0024] When coupled extreme scenarios are identified, the risk redundancy adjustment operator is activated to expand the constraint redundancy of flood control and ecological objectives and reduce the risk rate. Simultaneously, based on risk redundancy adjustments, the FLA emergency optimization mode is triggered to quickly generate optimization solutions adapted to extreme scenarios, including parameter adjustments and function optimizations. Parameter adjustments are underway: increasing population size to enhance search diversity; reducing the maximum number of iterations to accelerate computation; and adjusting the diffusion and penetration coefficients as needed to balance global search with local optimization. In the optimization of the degree function: the risk penalty coefficient is increased to strengthen risk constraints; additional reward items are set for core objectives in extreme scenarios, with the following formula: (32) in The fitness function is the core objective, and δ is the reward coefficient. Reduce the risk rates of flood control and ecological damage under extreme coupled conditions to avoid system failure.

[0025] In the preferred embodiment, step S5: The optimization method is based on the water level adjustment amount optimized by FLA. Based on the current actual water level of the reservoir, calculate the target water level range for future real-time or conventional models. Calculate the basic target water level at time k in the future The formula is: (33) in, To optimize the water level adjustment amount obtained in the i-th hour; The actual water level at the current time t; In determining the water level range, if the water level fluctuation range is set to float, then the target water level range at time k in the future will be: (34) At the same time, the target water level range must be within the upper and lower limits of the reservoir water level. If it falls below the lower limit, then adjust to ;like If it exceeds the upper limit, adjust to ; Considering the maximum variation in target water level between adjacent hours, to avoid drastic water level fluctuations, if If the difference is adjusted to the maximum change, the target water level range is recalculated.

[0026] The time-segmented water discharge baseline is based on the water discharge output optimized by FLA. Based on the water balance equation and constraints, the optimized water discharge rate is obtained. The water balance equation (20) must be satisfied; if If the water balance is not met, adjust according to the following formula: (35) in, The storage capacity constraint at time k+1; Adjusted water discharge volume The redistribution satisfies the formula: (36) in For the amount of water released for power generation; For flood discharge volume, For ecological water release; Ecological water release It should be greater than the lower limit of the ecological base current; the amount of water released for power generation The formula should be satisfied: (37) in, Let be the target power output at time k, ρ be the density of water, and g be the acceleration due to gravity. Let η be the water head at time k, and η be the power generation efficiency. The amount of water discharged should meet the following requirements. (38) The optimized control scheme should be output in a standardized form, including the scheduling period, target water level range, benchmark values ​​for water release in different time periods, risk rate assessment results, and operational precautions.

[0027] In the preferred embodiment, in step S6: The feedback data includes hourly data on the actual operation of the reservoir, which is compared with the optimized control scheme. Actual water level data is collected from the water level gauge at the top of the reservoir dam, averaging the water level for the current hour. and the target water level range contrast; The actual water discharge data is collected by the flow sensor at the outlet, which measures the total water discharge for the current hour. and the amount of water discharged Compare with the baseline value of water discharge; Actual ecological flow data is collected from downstream ecological flow monitoring stations to measure the actual ecological flow for the current hour. Compared with the ecological flow demand range; Actual power generation data is collected from the power generation system to measure the actual power generation for the current hour. Compared with the power generation target; Then, the deviation was calculated to obtain the water level deviation. Water discharge deviation and ecological flow deviation Based on the magnitude of the deviation, a graded correction strategy is adopted: When the water level deviates Water discharge deviation and ecological flow deviation When all values ​​are less than the first-level threshold, the deviation is small, and the current optimized control scheme remains unchanged. When the water level deviates Water discharge deviation and ecological flow deviation If any item exceeds the first-level threshold, the deviation is too large, and a linear correction method is used to adjust the operating parameters for the next hour. When the water level deviates Water discharge deviation and ecological flow deviation If any item exceeds the secondary threshold, the deviation is too large, triggering re-optimization. Return to step 2, input the current actual data, and re-execute the mapping, optimization, and risk management process to generate a new optimization control scheme. If the deviation is still too large after re-optimization, check the data acquisition equipment and model parameters, and perform manual intervention; By periodically performing closed-loop feedback corrections, the cumulative deviation of operating parameters is controlled within the ideal value.

[0028] This invention significantly improves real-time performance and adapts to fluctuating renewable energy demands through a collaborative logic of cross-scale fusion at the data layer, dynamic mapping at the modeling layer, dual-mode optimization at the algorithm layer, extreme risk control at the risk layer, and closed-loop correction at the implementation layer. Closed-loop feedback correction ensures dynamic adaptation of operating parameters to the actual scenario; it achieves multi-objective synergistic optimization with outstanding multi-objective balancing effects, safeguarding watershed ecological and flood control safety while improving scheduling economy; it operates stably under various coupled extreme scenarios, exhibits strong anti-interference capabilities, and is widely adaptable to reservoirs of different sizes, varying renewable energy integration ratios, and various hydrological scenarios; no additional hardware modifications are required, and ordinary industrial computers can support algorithm execution; it overcomes the bottlenecks of subjective weighting, blind risk control, and incompatibility between real-time performance and global optimization in existing technologies, effectively resolving the technical contradictions of multi-objective reservoir scheduling in renewable energy integration scenarios and providing technical support for the scientific scheduling and sustainable management of reservoirs. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of a real-time, high-frequency, cross-scale collaborative, dynamic adaptive FLA reservoir multi-objective scheduling method according to the present invention; Detailed Implementation Example 1 like Figure 1 As shown, a real-time, high-frequency, cross-scale collaborative, dynamic adaptive FLA reservoir multi-objective scheduling method is proposed. Taking cascade reservoirs as the specific application scenario, the core objectives cover flood control, power generation, ecological flow guarantee, hydro-solar complementary peak shaving, reservoirs and photovoltaic infrastructure; In the preferred scheme, step S1: First, high-frequency and low-frequency data are collected hourly, including: High-frequency data: One TDS-100H ultrasonic flow sensor is deployed in the main channel of the reservoir and at the inlets of the two main tributaries. The sensors are installed at the center line of the channel cross section and the instantaneous flow data of the three sensors are collected simultaneously. The arithmetic mean is taken as the inflow value for the current period. At the same time, the flow fluctuation amplitude and rate of change are recorded. The actual output of the photovoltaic array is collected through the SCADA system of the photovoltaic power station, and the irradiance and ambient temperature are collected simultaneously for subsequent anomaly detection. A Rio Grande acoustic Doppler current profiler was deployed downstream of the reservoir outlet to collect outflow data; ecological flow demand ranges were set in conjunction with ecological protection plans. Low-frequency data: The current flood control limit reservoir capacity is calculated based on the water level-reservoir capacity curve in the flood season scheduling plan and updated daily through the reservoir management system. The power generation target data is broken down into ten-day periods based on the annual power generation target, with the ten-day target updated on the first day of each ten-day period and broken down to the daily level. Static baseline data is verified by fitting historical data with field surveys at the end of each quarter based on the water level-output curve to reduce parameter errors.

[0030] The outlier removal operation groups the photovoltaic power output data by day, with a 30-day sliding window, for a total of 720 hourly data points; the power output mean μ within the window is calculated according to formula (1), and the standard deviation σ is calculated according to formula (2); Subsequently, the 3σ criterion was adopted. If an anomaly is detected, the value is removed.

[0031] Feature extraction was implemented using a Python environment. For high-frequency data, statistical features such as the mean and maximum value within a 12-hour sliding window were calculated for the inflow, as well as dynamic features such as the hourly rate of change and the 6-hour cumulative change, and the cross-correlation with photovoltaic output were also calculated. For low-frequency data, the trend slope was obtained by linear fitting of the power generation target for the first half of the month, and the flood control capacity and downstream flood level were transformed into utilization rate features. Finally, 22-dimensional high-frequency and 6-dimensional / 28-dimensional low-frequency feature vectors were generated using the Pandas and Scikit-learn libraries.

[0032] Cross-scale data fusion adopts a hierarchical dynamic factor model. Historical data from previous years are selected, and low-frequency data is linearly interpolated to the hourly level and then aligned. Three factor layers (input layer, trend, fluctuation, and synergy) and an output layer are set. The parameters are solved by 800 iterations using the maximum likelihood estimation method and the EM algorithm. If the root mean square error and the mean absolute percentage error calculated on the validation set are both less than 5%, the accuracy requirement is met. The 28-dimensional features of the current time period are input into the model, and cross-scale input data containing trend and fluctuation components are output according to formulas (3) and (4).

[0033] In the preferred embodiment, in step S2: When calculating cross-scale sequence parameters Inbound flow fluctuation range Calculated according to formula (5), the comparison period is selected as the previous 6 hours. Assuming that the average inflow rate during this period is 32200 m³ / s obtained by arithmetic average, and the inflow rate during the current period is 33400 m³ / s, therefore... To facilitate scenario assessment, further calculations are performed. ; Photovoltaic power output change rate According to formula (6), assuming the photovoltaic output is 85MW in the current period and 89MW in the previous hour, therefore The negative sign indicates that the photovoltaic output is declining; The ordinal parameter threshold is based on historical data statistics. The K-means clustering algorithm was used to cluster 34,560 hourly ordinal parameter data points from the past four years (48 months), with a cluster size of k=3. During the clustering process, the cluster centers stabilized after 100 iterations, assuming a low-fluctuation scenario. , Corresponding to conventional hydrological conditions and stable photovoltaic output; medium-fluctuation scenarios , This corresponds to mild hydrological fluctuations and general photovoltaic fluctuations; high-fluctuation scenarios. , This corresponds to the hydrological data related to heavy rainfall and the severe fluctuations in photovoltaic power. The minimum threshold for high volatility categories is taken as the warning threshold, i.e. Warning threshold 5% The warning threshold is 15%. The status is determined based on the current time period, and corresponding measures are taken.

[0034] The Bayesian dynamic factor model structure needs to match the requirements of multi-objective scheduling, where: The input variable X is the 28-dimensional cross-scale input data output from step S1, covering high-frequency dynamic features and low-frequency trend features; Four latent variables F are set, which correspond to the weighted driving factors of the four major objectives of flood control, power generation, ecology and peak regulation. Each latent variable can capture the dynamic changes of the corresponding objective. The output variable W is the weight vector of the four main objectives. The softmax function ensures that the sum of the weights is 1, where w1 is the flood control weight, w2 is the power generation weight, w3 is the ecological weight, and w4 is the peak shaving weight. The mathematical expression of the model is constructed according to formulas (7)(8)(9)(10). The model training was carried out by five senior scheduling experts who performed weight labeling. The labeling process adopted the Delphi method. First, the experts labeled independently, and the labeling was based on the hydrological scenario, the fluctuation of new energy sources and the priority of scheduling objectives. Then, three rounds of consultation meetings were organized to discuss the samples with large labeling differences, and finally a consistent labeling was formed.

[0035] The prior distribution factor loading matrix Λ is set to a normal prior N(0, 10I), where I is a 28th order identity matrix, and the variance setting of 10 ensures that the prior is sufficiently broad and does not dominate the data information; the autoregressive coefficient matrix Φ is set to a normal prior N(0, 0.1I), and the variance setting of 0.1 reflects the weak temporal correlation of latent variables, which is consistent with the short-term stability of the scheduling target weights; the observation error variance... State error variance Let the inverse Gamma prior be IG(2,1). This prior is a weak information prior, which is convenient for data-driven parameter estimation.

[0036] The MCMC sampling uses the Metropolis-Hastings algorithm with 10,000 samplings. The first 2,000 samplings are a warm-up period to eliminate the influence of initial values ​​on the sampling results. These samples are then discarded. The remaining 8,000 samples are used for parameter estimation. Convergence is determined by calculating the potential scaling factor. If the factor is ≤1.05, it indicates that the multi-chain sampling results are consistent, and convergence is determined.

[0037] Model validation uses confusion matrix and weight prediction error assessment: confusion matrix is ​​used to evaluate scene classification accuracy (such as the accuracy of predicting normal scene as normal). Weight prediction error uses mean absolute error (MAE) and is calculated according to formula (11). Data from August 2024 (744 data points) is selected as validation set. The calculated MAE is 2.5%≤3%, which indicates that the weights predicted by the model have a small deviation from the weights labeled by experts and can be used for real-time mapping.

[0038] Model validation employs a confusion matrix and weight prediction error evaluation, where: The confusion matrix is ​​used to evaluate the accuracy of scene classification. The weight prediction error is the mean absolute error (MAE), calculated according to formula (11). When MAE ≤ 3%, it indicates that the weights predicted by the model have a small deviation from the weights labeled by the experts, and can be used for real-time mapping. During weight prediction, the current cross-scale input data is input into the trained BDF model, which outputs the basic weight vector. Dynamic weight adjustment needs to be combined with the state of the order parameter: when the current state is normal, no additional adjustment is needed, and the output can be directly performed. If the subsequent period enters an abnormal state, such as when If the result is determined to be a single-factor high fluctuation, the peak-shaving weight is increased according to formula (15); the power generation weight is decreased according to formula (16); the flood control weight w1=0.3 and the ecological weight w3 are output as the final weight. The post-verification weights sum to 1, which satisfies the constraints. In special scenarios, the weights need to be further adjusted. For example, when a blue rainstorm warning is issued, an emergency flood control adjustment is initiated: the flood control weight w1 is increased to the preset upper limit of 0.45, and the peak shaving weight w4 is reduced to 0.1 to reduce the impact of peak shaving on flood control. After the adjustment, it is necessary to re-verify whether the flood risk rate is lower than the preset value of 3% to ensure that the weight adjustment is reasonable.

[0039] Flood control and ecological risk quantification transforms abstract risks into calculable probability indicators, among which flood control risk rate... The probability of the reservoir's outflow exceeding the downstream flood control safety discharge during the forecast period is defined as the probability of the outflow exceeding the downstream flood control safety discharge during the forecast period. The forecast period is set to the next 6 hours, and the selection of this duration needs to take into account the reservoir flood propagation time, so that flood control preparations can be made in advance during the forecast period. The inflow prediction adopts an LSTM model. The input of this model is the inflow data of the past 24 hours and the rainfall forecast data for the next 6 hours issued by the meteorological department. The output is the predicted inflow value for the next 6 hours and the 95% confidence interval. The confidence interval is obtained by fitting the normal distribution of the model prediction error. The flood risk rate is calculated using the Monte Carlo simulation method to handle uncertainties. First, 10,000 sets of inflow sequences are randomly sampled from a normal distribution N, with each set containing 6 hours of data. Then, for each sampled sequence, the current reservoir capacity is considered. Flood control weight in dynamic weighting The discharge flow sequence for the next 6 hours was calculated using the water balance equation (20). Finally, statistics were compiled from 10,000 samples to determine the discharge flow rate. Exceeding the downstream flood control safety discharge capacity Calculate the flood risk rate Does it meet safety requirements? Ecological risk rate Defined as the probability that the outflow from the reservoir during the forecast period will not meet the ecological flow pulse demand, the ecological flow demand range for the next 6 hours is determined according to the ecological protection plan. Using the Monte Carlo simulation method, which is the same as the flood control risk rate, the probability of outflow exceeding the reservoir flow is statistically analyzed in 10,000 sets of inflow sampling sequences. The number of intervals is used to calculate whether the ecological risk rate meets the ecological security requirements.

[0040] In the preferred scheme, step S3: Based on the order parameter state and dynamic weights, the current scheduling scenario is identified, and the improved Flood Algorithm (FLA) is designed with dual-mode adaptive logic to adapt to both real-time peak shaving and regular optimization scenarios. The initialization of FLA basic parameters needs to be combined with the requirements of the scheduling scenario. In the real-time peak shaving scenario, the solution vector dimension is set to 12, corresponding to the next 6 hours. Each hour contains 2 decision variables: water level adjustment amount. Water discharge ,Right now: ; Variable boundary constraints must balance engineering safety and operational flexibility. The range should be set based on the minimum adjustment accuracy of the reservoir gate. The maximum adjustment per hour should be less than this accuracy to avoid stress changes in the dam body caused by drastic fluctuations in water level. The lower limit is the ecological base flow, and the upper limit is the maximum discharge of the reservoir.

[0041] The criteria for scene recognition are divided into core criteria and auxiliary criteria. The core criterion is the peak-shaving weight output in step S2. Photovoltaic power output change rate The auxiliary indicator is the predicted fluctuation range of photovoltaic output over the next 6 hours, which is calculated by combining the photovoltaic power plant's SCADA system with weather forecasts. If the peak shaving weight is ≥30% or the absolute value of the photovoltaic output change rate is ≥15%, and the auxiliary indicator shows that the photovoltaic fluctuation range is ≥15% in the next 6 hours, then it is determined to be a real-time peak shaving scenario, triggering the FLA real-time response mode, and the system automatically loads the parameter configuration of the real-time mode.

[0042] The parameter settings for real-time response mode need to balance convergence speed and solution set quality. Too few populations, such as 20, result in fast convergence but insufficient solution set diversity; too many, such as 40, result in good diversity but excessive convergence time. A population of 30 can converge within the receiving time and the solution set HV overvolume ≥ 0.75; maximum number of iterations T = 300; diffusion coefficient α = 0.75, increasing the diffusion coefficient can expand the initial search range and quickly approach the peak-related solution space; penetration coefficient β = 0.2, weakening local search can reduce computation time and prioritize convergence speed; crossover probability... A higher crossover probability can accelerate the spread of superior genes in a population; mutation probability A lower mutation probability can prevent the population from becoming homogenized too early, without increasing the amount of computation.

[0043] The fitness function is constructed according to formula (23). γ should be determined through testing. If it is too small, the risk constraint is insufficient; if it is too large, excessive penalty leads to target imbalance. The calculation of the fitness of each target needs to be combined with the actual engineering situation to obtain a large value. Flood control target adaptability Power generation target adaptability Ecological target fitness Peak-shaving target fitness reflects the complementarity between fluctuations in renewable energy output and fluctuations in reservoir power output, serving as the target; the final fitness... It comprehensively reflects the balancing effect of multiple objectives.

[0044] In the algorithm iteration, 30 initial solution vectors are first randomly generated within the variable boundary. The generation of each solution vector must satisfy the boundary constraints. Then, a global search is performed on each solution vector according to formula (21) to generate a new solution. The boundary needs to be rechecked for the new solution. If it exceeds the boundary value, it is truncated to the boundary value. Next, local optimization is performed on the solution vectors with the top 30% fitness according to formula (22) to refine the surrounding area of ​​the high-quality solution and improve the accuracy of the solution. Then, single-point crossover and polynomial mutation are used to generate offspring solutions. The crossover point is randomly selected to correspond to the water release volume of the corresponding hour. The part after the crossover point is swapped to generate offspring. Polynomial mutation Adjustments are made; finally, a convergence check is performed. If the HV hypervolume change of the non-dominated solution set is less than 10⁻³ for 20 consecutive iterations, or the maximum number of iterations (300) is reached, iteration is stopped, and the optimal solution is output. Its adaptability This is the optimal value in the current population.

[0045] Optimization performance evaluation requires real-time monitoring of solution set quality, using three metrics: HV hypervolume, GD convergence distance, and SP distribution. The HV hypervolume is calculated according to formula (24). If it is higher than the initial population, it indicates that the convergence and distribution of the solution set are good. The convergence distance of GD is calculated according to formula (25). If the current GD < 0.15, it indicates that the solution set is close to the true Pareto front, that is, the theoretical optimal solution set obtained by a large number of iterations. The SP distribution is calculated according to formula (26). If the current SP < 0.08, it indicates that the solution set is uniformly distributed and there is no clustering. If all three indicators meet the performance requirements of the real-time peak shaving scenario, then the output optimal solution has engineering applicability.

[0046] In the preferred embodiment, step S4: Identify coupled extreme scenarios such as heavy rain and insufficient inflow, and make up for the risk coverage gap of conventional dual modes through risk redundancy adjustment and FLA emergency optimization to ensure that flood control and ecological targets are not exceeded under extreme conditions; Among them, the identification of coupled extreme scenarios requires combining the state of the sequence parameters, the risk rate threshold, and logical judgment. First, the real-time sequence parameters and risk rate output in step S2 are obtained, along with the fluctuation range of the inbound traffic. Photovoltaic power output change rate Flood risk rate Ecological risk rate Then, obtain the current ecosystem traffic demand and inbound traffic level, and define coupled extreme scenarios based on extreme scenarios. or or The current situation is determined to be an extreme coupled scenario, and the risk management process needs to be initiated.

[0047] Uncertainty quantification is a prerequisite for risk management in extreme scenarios. A Box-based polyhedral ensemble is used to process hydrological forecast uncertainty. The Box set describes the range of prediction errors, while the polyhedral ensemble describes the probability distribution characteristics of the errors, thus comprehensively quantifying uncertainty. First, a Box set is constructed. Assuming the predicted inflow to the reservoir is I = 42000 m³ / s, and the maximum error in the reservoir inflow forecast is 10%, therefore… m³ / s, according to formula (27) Box={I∈R | 42000-4200≤I≤42000+4200}={37800≤I≤46200}, which covers more than 95% of the historical forecast error; Next, a polyhedral set is constructed, based on the prediction error data. The fitting error follows a normal distribution N(0,σ²), where... m³ / s, where 1.96 corresponds to the 95% confidence interval. According to formula (28) Poly={I∈R | (I-42000)² / (2142.86)²≤3.841}, where χ²(1,0.95)=3.841 is the chi-square distribution critical value with 1 degree of freedom and 95% confidence level. The calculated value is 37800≤I≤46200m³ / s, which is consistent with the Box set, indicating that the forecast error conforms to the normal distribution and the uncertainty quantification result is reliable. Uncertainty propagation calculations require the use of the water balance equation. First, the uncertainty set of the inflow is obtained: U = Box∩Poly = {37800≤I≤46200} m³ / s. Then, the outflow is calculated based on formula (30). Finally, the maximum possible value of the flood risk rate under extreme scenarios is calculated according to formula (31). By sampling 10,000 inbound flow sequences in Monte Carlo, the number of sequences that meet the conditions is counted to obtain the risk rate. If it is greater than 3%, it needs to be reduced to below the threshold through risk redundancy adjustment.

[0048] The risk redundancy adjustment first adjusts the flood control limit water level and flood discharge flow, lowering the flood control limit water level by a margin determined based on historical rainstorm data to further reduce downstream flood control pressure. Then, the target weights are adjusted to prioritize flood control and peak shaving targets. The flood control weight w1 is increased to 0.45 (the preset upper limit), the peak shaving weight w4 is increased to 0.3, the power generation weight w2 is reduced to 0.15 to reduce the water consumption for power generation from occupying flood control reservoir capacity, and the ecological weight w3 is reduced to 0.1 to ensure ecological base flow. After the adjustment, the weights must be 1 to cope with the needs of extreme scenarios.

[0049] FLA emergency optimization requires rapidly generating optimization schemes adapted to extreme scenarios based on risk redundancy adjustments; parameter adjustments need to increase the population size to cover a wider solution space and capture the optimal solution under extreme scenarios; the maximum number of iterations should be increased to ensure sufficient search while keeping the time consumption within an acceptable range; the diffusion coefficient should be reduced to balance global search and local optimization; the penetration coefficient should be increased to strengthen local search and adapt to the adjusted constraint boundary; the crossover probability and mutation probability should be adjusted as needed to maintain the efficiency of high-quality gene propagation.

[0050] Fitness function optimization needs to strengthen risk constraints and core objectives, increase the risk penalty coefficient to increase the penalty for exceeding the risk limit; add a core objective reward item, and guide the algorithm to prioritize searching for solutions with excellent core objectives according to formula (32).

[0051] The algorithm iterates until HV≥0.80 and GD≤0.08, then outputs the optimal solution, which satisfies the requirements of risk control and multi-objective balance in extreme scenarios.

[0052] In the preferred embodiment, step S5: The abstract solution vector obtained by FLA optimization is transformed into a specific and operable reservoir scheduling and control scheme, thereby improving the engineering practicality of the scheme; The calculation of the target water level range needs to combine the current actual water level with the water level adjustment amount optimized by FLA. First, the actual water level at the current time t needs to be obtained. This data is collected in real time by a precision water level gauge on the dam crest, recorded every 10 minutes, and the average value of the first 10 minutes of each hour is taken as the current water level; the water level adjustment amount for this period is then optimized and output by FLA. Substitute into formula (33) to calculate the basic target water level This serves as the ideal water level for that specific time period.

[0053] To accommodate the flexibility of engineering operations, the water level fluctuation range is set to float=0.1m. This range should be set according to the minimum adjustment accuracy of the reservoir gate. A fluctuation of 0.1m allows for fine-tuning on-site according to the actual situation, while avoiding operational chaos caused by excessive fluctuations. At the same time, after calculating the target water level range according to formula (34), it is verified whether the range is within the upper and lower limits of the reservoir water level. It is also necessary to check whether the water level change range of adjacent hours is less than the maximum allowable change per hour. If the change range is less than the maximum allowable change, the base water level needs to be adjusted to the water level of the previous period before recalculating the range.

[0054] The calculation of the benchmark value of water release in different time periods needs to first verify the feasibility of the FLA optimization results before making reasonable allocations. First, verify whether the deviation between the FLA optimized water release and the target reservoir capacity is <0.03 according to formula (20). If the water balance constraint is met, then the optimized water release is feasible.

[0055] The allocation of water release volume first determines the ecological water release volume. Whether it exceeds the lower limit of ecological base flow, and verified by downstream ecological monitoring data; then calculate the power generation discharge volume according to formula (37). Finally, calculate the flood discharge volume according to formula (38). Check whether the inflow rate is less than the reservoir's inflow rate to avoid exceeding the upper limit of the water level.

[0056] The output of the control plan needs to take into account both manual decision-making and automatic execution. First, it is output to the dispatch center decision-making system in the form of text instructions. The content includes the specific time period, the target water level range of the reservoir, the total water release volume including the ecological water release volume, the power generation water release volume, and the flood discharge water release volume; whether all safety constraints are met, and whether emergency measures need to be activated. Furthermore, it simultaneously generates PLC control commands that can be recognized by the field control system. The commands adopt the Modbus protocol and include gate number, target opening degree, and adjustment rate. The commands are transmitted to the reservoir's field control unit via a 5G private network to ensure timely execution.

[0057] In the preferred embodiment, step S6: Construct a closed-loop mechanism of optimization, execution, feedback, and correction to eliminate the cumulative deviation between operating parameters and actual hydrological scenarios in real time, maintain the continuous effectiveness of the scheduling plan, and avoid the amplification of deviations; Feedback data collection needs to be carried out synchronously at the end of each hour, covering four core data categories: water level, water release, ecological flow, and power generation. Among them, the actual water level data is collected by a precision water level gauge on the top of the dam, recording the instantaneous water level value every 6 minutes, and taking the arithmetic mean to obtain the actual water level. The water level deviation is calculated by comparing it with the midpoint of the target water level range. ; The actual water release data is collected by flow sensors at the reservoir outlet, recording the hourly discharge flow of the ecological gate, power generation gate, and flood discharge gate, including ecological water release, power generation water release, and flood discharge water release. The total water release is obtained by adding them together. The relative deviation is calculated by comparing it with the baseline value of water discharge. ; Actual ecological flow data is collected by the ADCP at the downstream hydrological station, recording the average ecological flow for that period. The ecological flow deviation is calculated by comparing it with the midpoint of the ecological flow demand range. .

[0058] The actual power generation data is collected by the power generation monitoring system of the reservoir power station, which records the total power generation for the current period. Compare the deviation rate with the power generation target, determine the type of deviation, and whether it is within an acceptable range; The determination of its deviation level must be based on a preset threshold, and the threshold setting is based on the statistical analysis of historical operation data of the reservoir: The first-level threshold is considered a slight deviation, with water level deviation ≤0.05m, relative deviation of water release ≤0.5%, and ecological flow deviation ≤50m³ / s. Deviations within this range have little impact on the scheduling target and do not require adjustment. The secondary threshold is considered a moderate deviation: 0.05m < water level deviation ≤ 0.1m, 0.5% < water discharge deviation ≤ 1.2%, and 50m³ / s < ecological flow deviation ≤ 100m³ / s. It needs to be adjusted through linear correction. Level 3 thresholds are considered serious deviations, with water level deviation > 0.1m, water discharge deviation > 1.2%, and ecological flow deviation > 100m³ / s. It is necessary to return to step S2 for re-optimization. The adjustment coefficient for linear correction needs to be determined through PID control simulation to ensure stable correction effect without overshoot; the water level correction coefficient is obtained through simulation testing; the discharge volume correction coefficient needs to ensure that it can quickly approach the target value without over- or under-oscillation; the ecological flow correction coefficient is adjusted according to the ecological conditions. The corrected solution needs to be sent to the on-site execution system in the next time period, and data should be collected again to verify the correction effect, ensuring that the cumulative deviation of the operating parameters is ≤0.5% to avoid the amplification of deviations during long-term operation; if a moderate deviation occurs for 3 consecutive hours, the correction coefficient should be automatically increased to enhance the correction effect; if a serious deviation still occurs after correction, the accuracy of the data acquisition equipment should be checked, and manual intervention should be triggered if necessary to ensure the effective operation of the closed-loop mechanism.

[0059] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A real-time, high-frequency, cross-scale collaborative, dynamic adaptive FLA reservoir multi-objective scheduling method, characterized in that, Includes the following steps: S1. Collect multi-dimensional and multi-frequency scheduling basic data, remove outliers and extract features from the basic data, and then integrate it into cross-scale input data containing high-frequency dynamic features and low-frequency trend features through a cross-scale fusion model. S2. Identify cross-scale order parameters and determine their state. Construct a cross-scale mapping model based on the Bayesian dynamic factor algorithm, generate multi-objective dynamic weights, and combine probabilistic statistical methods to quantify flood risk rate and ecological risk rate. S3. Based on the state of the order parameter and the dynamic weight to identify the scheduling scenario, the flood algorithm (FLA) is improved to design a dual-mode adaptive logic, which is adapted to the real-time peak shaving scenario and the conventional optimization scenario respectively. The optimal solution vector is output through iterative optimization. S4. Based on the identification of extreme scenarios coupled with the state of the ordinal parameter and the risk rate threshold, the uncertainty of hydrological forecasting is quantified by set quantification, and an optimization scheme adapted to extreme scenarios is generated through risk redundancy adjustment and FLA emergency optimization. S5. Combining the current actual state of the reservoir with the optimal solution vector, calculate the target water level range and the benchmark value of water release in different time periods, and output an operable reservoir scheduling and control scheme. S6. Collect actual reservoir operation data, calculate the deviation from the scheduling and control scheme, implement graded correction strategies according to the deviation level, and form a closed-loop feedback mechanism.

2. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, the steps are as follows: In S1: High-frequency data includes inbound flow data, new energy output data, and ecological flow pulse demand data; Low-frequency data includes flood control capacity data, long-term power generation target data, and static basic data of reservoirs; Outlier removal is achieved through a sliding window combined with statistical criteria; Feature extraction includes calculating the statistical features, dynamic features, and correlation features of the data; Cross-scale fusion is accomplished by a hierarchical dynamic factor model that includes an input layer, a factor layer, and an output layer.

3. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 2, characterized in that, Step S1 also includes: S11. When training the hierarchical dynamic factor model, select historical data and upscale the low-frequency data to the same frequency as the high-frequency data to form an aligned training set. The model parameters are estimated using the maximum likelihood estimation method, and the optimal solution of the parameters is solved by the expectation-maximization algorithm until the parameter update magnitude is less than the preset condition. The model fusion effect is verified based on the root mean square error and the mean absolute percentage error. If the preset accuracy is not met, the number of factors is adjusted and the model is retrained.

4. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, In step S2: The cross-scale sequence parameters include the fluctuation range of inflow and the rate of change of renewable energy output, among which: The fluctuation range of inbound flow is defined as the absolute value of the difference between the current hourly inbound flow and the average inbound flow over a set period of time. The rate of change of renewable energy output is defined as the relative change in renewable energy output in the current hour compared to the renewable energy output over a previously set time period. Clustering algorithms are used to classify historical sequence parameter data into fluctuation levels, determine the warning thresholds corresponding to each level, and combine the real-time sequence parameter values ​​to determine whether the status is normal, single-factor high fluctuation, or double-factor high fluctuation.

5. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, In step S2: The flood risk rate is defined as the probability that the reservoir discharge will exceed the downstream flood control safety discharge during the forecast period. The ecological risk rate is defined as the probability that the outflow from the reservoir during the forecast period will not meet the ecological flow pulse demand. A long short-term memory network model is used to predict the inflow and corresponding confidence interval for a set period of time in the future. Multiple inflow sequences are generated by sampling from the probability distribution of the flow prediction using the Monte Carlo simulation method. The discharge flow or outflow corresponding to each sequence is calculated by combining the water balance equation. The percentage of the number of sampling groups that meet the risk conditions is counted to obtain the risk rate.

6. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, Step S2 also includes: S21. The input variables of the Bayesian dynamic factor model are the cross-scale input data output in step S1. The latent variables correspond to the weight driving factors of flood control, power generation, ecology and peak regulation. The output variables are multi-objective weight vectors and satisfy the weight sum of 1. The model parameters are solved by MCMC sampling method. A warm-up period is set during the sampling process to eliminate the influence of initial values. The sampling convergence is judged based on the potential scale reduction factor. The basic weights of the model output are adjusted according to the state of the order parameter. In special scenarios, the weights of the core target are prioritized to increase and ensure that the sum of the weights of all targets is still 1 after adjustment.

7. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, In step S3: The core indicators for scene recognition are the peak shaving weight and the rate of change in new energy output output output in step S2. The auxiliary indicator is the fluctuation range of the predicted output of new energy sources within a set future time period; In real-time peak shaving scenarios, the population size, number of iterations, diffusion coefficient, penetration coefficient, and crossover probability of the flood algorithm are adjusted to enhance its fast convergence capability. In conventional optimization scenarios, the crossover and mutation probabilities are dynamically adjusted based on the HV hypervolume index to balance the accuracy of global search and local optimization. The algorithm's performance is optimized in real time by using HV hypervolume, GD convergence distance, and SP distribution index to maintain the convergence, distribution, and uniformity of the output solution set.

8. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, In step S4: When identifying coupled extreme scenarios, it is necessary to combine the order parameter status and risk rate threshold output in step S2 with the current ecological traffic demand and inbound traffic level for logical judgment. The uncertainty quantification of hydrological forecasts uses the intersection of Box sets and polyhedral sets. The Box set describes the range of prediction error, while the polyhedral set describes the probability distribution characteristics of the error. Risk redundancy adjustment includes adjusting the constraint boundaries of flood control or ecological objectives and optimizing objective weights. FLA emergency optimization strengthens risk constraints by adjusting parameters such as population size and number of iterations, and adding a core objective reward item to the fitness function.

9. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, Its characteristic is that step S5 middle: When calculating the target water level range, the basic target water level is obtained based on the current actual water level of the reservoir and the water level adjustment amount output in step S3. The water level fluctuation range is set to form the target water level range. At the same time, it is checked whether the range is within the upper and lower limits of the reservoir water level and the change of the basic target water level in adjacent hours is controlled to not exceed the set threshold. When calculating the benchmark value of water release volume in different time periods, first verify whether the water release volume output in step S3 meets the water balance constraint. If it does not meet the constraint, readjust it based on the reservoir capacity constraint. The allocation of water release follows the principle of prioritizing ecology, followed by power generation, and supplementing flood discharge. It ensures that the ecological water release is not lower than the lower limit of the ecological base flow, the power generation water release is calculated based on the power generation target output, water head, and power generation efficiency, and the flood discharge water release is the remaining amount after deducting the ecological water release and the power generation water release from the total water release.

10. The real-time high-frequency cross-scale collaborative dynamic adaptive FLA reservoir multi-objective scheduling method according to claim 1, characterized in that, In step S6: The actual operational data collected includes actual water level, actual total water release and individual water release, actual ecological flow, and actual power generation. The calculated deviations include water level deviation, relative water release deviation, and ecological flow deviation. The graded correction strategy is as follows: If all deviations are less than the first-level threshold, maintain the current scheme; If any deviation is higher than the first-level threshold but lower than the second-level threshold, a linear correction method is used to adjust the operating parameters for the next time period. If any deviation exceeds the secondary threshold, return to step S2 to re-execute cross-scale mapping, risk quantification, and subsequent processes; If a moderate deviation occurs repeatedly, increase the linear correction coefficient. If a severe deviation still occurs after re-optimization, check the data acquisition equipment and model parameters and trigger manual intervention.

Images