Intelligent decision-making method for flood control dispatching of basin reservoir group system

Abstract

The invention provides an intelligent decision-making method for flood control scheduling of a basin reservoir group system. The method comprises the steps of advanced judgment of basin flood control risks, recognition of similar flood scenes, preliminary selection of flood control scheduling schemes, iterative optimization of the flood control scheduling schemes and risk feedback and scheme iterative optimization of the reservoir group system. According to the method, the problem of low efficiency of real-time flood control scheme calculation of an existing basin complex reservoir group is solved, and rapid flood control scheduling intelligent decision-making based on historical scene information and a loop optimization algorithm is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of water resources management, and particularly relates to an intelligent decision-making method for flood control scheduling of a basin reservoir group system. Background Art

[0002] Flood control scheduling of a basin reservoir group is an important means for basin water resources management, disaster prevention and reduction, and protection of people's lives and property. Reservoir flood control scheduling can be divided into conventional scheduling and optimized scheduling. Conventional scheduling adjusts and stores floods according to real-time incoming water conditions with a fixed discharge scheme; optimized scheduling uses incoming water forecast information and adopts algorithms to formulate a scheduling scheme that meets flood control objectives. The dynamic programming algorithm is a classic optimized algorithm for single reservoir flood control scheduling. It divides the scheduling period into multiple stages and saves the optimal strategy for each stage, thus obtaining a higher solution efficiency than the traversal method. However, a large number of intermediate variables will be generated during the calculation process of the dynamic programming algorithm, and the "curse of dimensionality" problem will occur when dealing with complex joint scheduling tasks of reservoir groups. Intelligent optimization algorithms represented by genetic algorithms can avoid the "curse of dimensionality" problem and have been widely used in the formulation of reservoir group scheduling schemes. However, the search space of the algorithm is large, and the time spent on solving is relatively long, making it difficult to meet the requirements of digital twin basins in terms of real-time performance and accuracy.

[0003] With the advancement of urbanization and water conservancy project construction, there are many key flood control nodes in the basin, and the requirements of each flood control node are different. Therefore, the formulation of flood control scheduling schemes is a multi-objective decision-making problem. The current decision-making methods mainly give decision-makers' preference information in advance and then select a scheme from the set of schemes. The decision-making results are significantly affected by decision-makers' subjective preferences, and there is a lack of the ability to timely feedback risks according to the development of the basin flood control situation and adjust the decision-making scheme.

[0004] Therefore, it is urgent to innovate intelligent methods for flood control scheduling of basin reservoir groups, improve the intelligent level of decision support and decision-making efficiency, and provide technical support for basin flood control and disaster reduction operations. Summary of the Invention

[0005] To overcome the problems of the prior art, the present invention proposes an intelligent decision-making method for flood control scheduling of a basin reservoir group system. The method uses a deep learning algorithm to predict the flood control risks of the basin, determines an initial feasible scheduling scheme based on similar historical scenarios, historical scheduling schemes, and combined with risk prediction information, uses an iterative optimization algorithm to optimize the scheduling scheme to eliminate the influence of flood scenario differences on the reservoir scheduling process, and finally uses a dynamic Bayesian network to infer the potential flood control risks of the basin under the joint action of scenarios and schemes, forming a cyclic iterative optimization process chain of "risk inference - scheme optimization", and finally obtaining a scheduling scheme that can coordinate the flood risks of the entire basin.

[0006] The object of the present invention is achieved as follows:

[0007] The present invention provides an intelligent decision-making method for flood control operation of a basin reservoir group system, including the following steps:

[0008] Step 1, early discrimination of basin flood control risks:

[0009] First, determine the scope of the flood control target basin and identify the key flood control nodes within the flood control target basin;

[0010] Then, by matching the relationship between the historical data of cumulative precipitation forecast in the flood control target basin and the actual disaster situation of key flood control nodes, construct a historical flood risk event dataset; construct a deep residual network model, and train and verify the model on the flood risk event dataset;

[0011] Input the real-time cumulative precipitation forecast data of the flood control target basin into the above-trained and verified model to complete the prediction of future potential flood control risks.

[0012] Step 2, identification of similar flood scenarios:

[0013] Collect the measured data of historical flood sites in the flood control target basin, establish a historical flood scenario database, and obtain historical generalized floods through mixed Gamma function fitting and normalization processing;

[0014] Select the Pearson correlation coefficient ρ, Euclidean distance ρ eu , dynamic time warping DTW, and mean absolute error MAE as flood similarity evaluation indicators, calculate the current forecast flood and the current generalized flood; then, calculate the similarity between the current generalized flood and the historical generalized floods, and match the historical flood event with the highest similarity to the current forecast flood.

[0015] Step 3, preliminary selection of flood control operation plans:

[0016] Use the multi-objective optimization algorithm (NSGA-Ⅲ) to offline solve the feasible operation plan sets of each reservoir in the reservoir group for each historical flood event, and store them in the historical flood scenario database to obtain a dispatching pre-plan library;

[0017] When a new flood event is about to occur, first use the method in Step 1 to early discriminate the basin flood control risks, then match the historical flood event with the highest similarity to the current forecast flood according to the method in Step 2, and at the same time match the feasible operation plan sets; determine the subjective weight of the decision according to the risk prediction situation of each reservoir itself or downstream of the reservoir, and use the augmented scale function decomposition method (ASF) to select the flood control operation plan that best meets the decision-maker's preference from the dispatching pre-plan library to complete the preliminary selection of the flood control operation plan and obtain the initial flood control operation plan.

[0018] Step 4, iterative optimization of flood control operation plans:

[0019] After determining the initial flood control operation plan, for each reservoir in the flood control target basin, input the current forecast inflow hydrograph to simulate the water level process during the reservoir operation period, and compare it with the historical reservoir water level process under historical flood events to obtain the differences in the highest water level, the lowest water level, and the end-of-period water level;

[0020] Use the operation plan cyclic optimization algorithm to adjust the initial flood control operation plan to make the highest water level, the lowest water level, and the end-of-period water level of the reservoir water level process under the action of the current forecast inflow hydrograph consistent with the historical process, and obtain the optimized flood control operation plan.

[0021] Step 5, risk feedback of the reservoir group system and iterative optimization of the plan:

[0022] Taking all reservoirs and key flood control sections in the flood control target basin as flood control key nodes, construct a dynamic Bayesian network model, and use the flood control key nodes as the nodes of the dynamic Bayesian network; taking the historical scenarios including historical flood events, feasible operation plans, reservoir operation water levels, and downstream flood control section hydrographs in Step 3 as input variables, and use the Monte Carlo simulation method to train the dynamic Bayesian network model;

[0023] In real-time flood events, taking the current forecast hydrograph, the optimized flood control operation plan obtained in Step 4, and the downstream section hydrograph as inputs, use the dynamic Bayesian network model to infer the risk probabilities of each flood control node in the basin;

[0024] Set the safety threshold of acceptable risk. If the inference result of the dynamic Bayesian network model shows that the risk probability of a certain node exceeds the safety threshold, it is judged that there is a flood control risk at this node, then formulate new reservoir water level and flow boundaries, and use the operation plan cyclic optimization algorithm to re-optimize the operation plans of each reservoir; after optimization, substitute it into the dynamic Bayesian network model again for risk inference, and repeat the above steps until the potential risks of all nodes are within the acceptable range, so as to complete the intelligent decision-making of flood control operation for the basin reservoir group system.

[0025] The advantages and beneficial effects of the present invention are:

[0026] Compared with the existing "forecast - scheduling - decision-making" process chain of basin reservoir groups, the intelligent decision-making method for flood control scheduling of the basin reservoir group system described in the present invention avoids directly using real-time forecast information to run the optimal scheduling model, and retrieving the scheduling plan from the historical scenario library saves the time cost of plan calculation. The risk prediction technology based on deep learning methods provides a basis for multi-objective decision-making, reduces the influence of subjective preferences in the decision-making process, and makes the decision result more in line with the real-time flood control situation of the basin. The cyclic iterative optimization algorithm can optimize the scheduling plan according to the differences between historical scenarios and the current scenario, and the risk feedback model based on the dynamic Bayesian network can real-time feedback the potential risks of the basin under the action of the scheduling plan, and support the cyclic iterative optimization of the plan. The present invention improves the decision-making efficiency and intelligent level of basin flood control scheduling compared with the prior art, and is more in line with the requirements of the construction of digital twin basins. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described below with reference to the drawings and embodiments.

[0028] Figure 1 is a schematic flowchart of the intelligent decision-making method for flood control scheduling of the basin reservoir group system according to an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of early prediction of basin flood control risks based on Resnet-50;

[0030] Figure 3 is a flowchart of the real-time risk feedback and scheduling plan optimization framework based on the dynamic Bayesian network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] Embodiment 1:

[0032] As Figure 1 shown, this embodiment provides an intelligent decision-making method for flood control scheduling of a basin reservoir group system, including early discrimination of basin flood control risks ( Figure 2 ), identification of similar flood scenarios, preliminary selection of flood control scheduling plans, iterative optimization of flood control scheduling plans, and risk feedback and plan iterative optimization of the reservoir group system ( Figure 3 ), and the specific method is as follows:

[0033] Step 1, early discrimination of basin flood control risks:

[0034] S11, analyze the historical raster data of basin cumulative precipitation forecasts:

[0035] Determine the scope of the flood control target basin and identify the key flood control nodes within the flood control target basin. The key flood control nodes include densely populated cities built along water and medium and small reservoirs prone to flood control risks.

[0036] Obtain and analyze the historical data of precipitation forecast products within the basin, including the historical data of cumulative precipitation forecasts for 3 hours, 6 hours, and 12 hours, and process them into two-dimensional raster images in.tif format, that is, the historical precipitation forecast data map.

[0037] In this embodiment, the precipitation forecast product can be selected from the precipitation forecast products of the Global Forecast System (GFS) or the European Centre for Medium-Range Weather Forecasts (ECMWF).

[0038] S12. Determine the key flood control nodes within the basin and determine the risk quantification criteria:

[0039] Take the locations of key flood control sections and water projects within the flood control target basin as the key flood control nodes, and determine the safety criteria for each key flood control node. The safety criteria include the safe flow rate of the river section, the safe water level, and the safe water level of the reservoir.

[0040] S13. Match the historical precipitation forecast data map with each risk category according to time:

[0041] First, collect the measured data of historical flood sites within the flood control target basin, and correspond the forecast product data with the measured data of historical flood events one by one. For example, the forecast cumulative precipitation from 1 hour to 7 hours corresponds to the measured flow rate or water level at 7 hours.

[0042] Classify the risk scenarios based on whether there is a risk at the key flood control nodes during historical flood events. If there is a flood risk exceeding the standard at node i during a historical flood event, it is defined as category i-0; if there are flood risks exceeding the standard at both node i and node j, it is defined as category ij-0; if no node has a flood risk exceeding the standard, it is defined as category none.

[0043] Match the processed forecast data map in S11 with historical flood events according to the occurrence time and put it into each classification to construct the historical flood risk data set.

[0044] S14. Use the historical flood risk data set to train the deep residual network model:

[0045] Construct the deep residual network model Resnet-50, and divide the historical flood risk data set into a training set and a validation set according to a ratio of 7:3; train the Resnet-50 network on the training set and optimize the hyperparameters to ensure that the prediction accuracy of the model on the validation set reaches more than 95%.

[0046] S15. Input real-time forecast data into the model and predict future flood control risks:

[0047] Obtain the real-time cumulative precipitation forecast data of the flood control target basin, process it into a two-dimensional raster image in.tif format according to S11, and input it into the trained deep residual network model Resnet-50 to predict the potential flood control risks of key flood control nodes in the basin.

[0048] Step 2, identification of similar flood scenarios:

[0049] S21, establish a historical flood scenario database:

[0050] Determine the fixed flood forecast lead time as the flood duration, and store the historical inflow process of each reservoir and the historical interval inflow process from the reservoir to the downstream section together after matching them according to time, so as to establish a historical flood scenario database.

[0051] S22, generalization of flood events:

[0052] First, perform normalization processing on flood events. The calculation formula for the normalization method is:

[0053]

[0054] In the formula: q(t) is the normalized flow value; Q(t) is the original flow value, m 3 / s; Q max and Q min are the maximum and minimum values of the original flow series respectively, m 3 / s; t’ is the normalized time; t is the original time, h; LT is the lead time, h.

[0055] Then, based on the characteristic differences between the rising and falling stages of the flood process, fit the flood process with a mixed Gamma function to obtain a generalized flood process;

[0056] The formula for the mixed Gamma function is:

[0057]

[0058] In the formula: k 1 、k 2 are normalization constants to ensure the continuity of the function processes in the two stages; t’ rise and t’ fall are the time scales of the rising and falling stages respectively; t’ peak is the time when the flood peak appears; α 1 and α 2 are the shape parameters of the Gamma function in the rising and falling stages respectively, Γ() is the Gamma function; e is the natural constant; y() is the generalized flood hydrograph.

[0059] The parameters in the mixed Gamma function are calibrated using the NSGA-Ⅲ algorithm, and three objective functions are set:

[0060] F 1 = min{q fit - q obs}

[0061] F 2 = min{w fit - w obs}

[0062] F 3 = max{NSE(fit, obs)}

[0063] In the formula: F 1 , F 2 and F 3 are the three calibration objectives respectively; q fit and q obs are the generalized flood peak size after normalization and the measured flood peak size after normalization respectively; w fit and w obs are the generalized flood volume after normalization and the measured flood volume after normalization respectively; fit and obs are the generalized process after normalization and the measured process after normalization respectively; NSE() is the Nash-Sutcliffe coefficient.

[0064] The generalized flood process and the original flood process are jointly stored in the historical flood scenario database.

[0065] S23, construct a similarity evaluation index system:

[0066] Select the Pearson correlation coefficient ρ, Euclidean distance ρ eu , dynamic time warping DTW, and mean absolute error MAE as evaluation indicators, and the calculation formulas are as follows respectively:

[0067]

[0068]

[0069] In the formula: f and his are the generalized processes of the predicted flow process and the historical flow process respectively; cov() is the covariance, σ f and σ his are the standard deviations of the predicted flow process and the historical flow process respectively; N is the total number of time periods; q f (t') and q his (t') are the generalized flow values of the predicted process and the historical process at time t' respectively; DTW() is the dynamic time warping algorithm, K is the total number of paths, and w k is the kth sub-path.

[0070] S24, Identification of similar flood scenarios:

[0071] Based on the forecast precipitation process, use a hydrological model (including a rainfall-runoff forecast scheme or a watershed hydrological model) to calculate the forecast flow process; calculate the current generalized flood using the method of S22, then use the index system of S23 to calculate the similarity index value between the current generalized flood and each historical generalized flood, and normalize the calculated index value; among them, the Pearson correlation coefficient is a positive index, and the Euclidean distance, dynamic time warping, and mean absolute error are negative indexes; the positive index remains unchanged, and the negative index takes the reverse value; assign a subjective weight W to each index value T , and obtain the similarity score;

[0072] Conduct a similarity search for flood events, that is, search in the historical flood scenario database for the historical flood generalization process that is most similar to the current forecast flood generalization process according to the similarity score from high to low. The similarity search is divided into the following 3 steps:

[0073] (a) Select 20 historical flood events with the highest similarity scores to form a flood event set A 1 ;

[0074] (b) In A 1 , eliminate the flood events with a relative peak flood error greater than 30% to obtain a flood event set A 2 ;

[0075] (c) In A 2 , eliminate the flood events with a relative flood volume error greater than 30% to obtain a flood event set A 3 .

[0076] Finally, select the historical generalized flood with the highest similarity score in A 3 , and after performing anti-normalization calculation on this process, obtain the historical flood event that is most similar to the current forecast flood event.

[0077] Step 3, Preliminary selection of flood control operation plans:

[0078] S31, Offline calculation of the multi-objective optimal flood control operation plans for each reservoir under historical flood event scenarios:

[0079] Use a multi-objective optimization algorithm (NSGA-Ⅲ) to calculate the multi-objective optimal operation plans for each reservoir in the reservoir group for each historical inflow flood event to obtain a set of feasible operation plans; match the set of feasible operation plans with the flood events and store them together in the historical flood scenario database to form a dispatching pre-plan database.

[0080] S32, Determine the decision-making preference weights for each reservoir:

[0081] Perform early discrimination on the flood control risks of the river basin according to the method in Step 1, and then match the historical flood event with the highest similarity to the current forecast flood in history according to the method in Step 2, and at the same time match the set of feasible dispatching schemes;

[0082] Determine the subjective weight of the decision according to the risk pre-judgment of each reservoir itself or downstream of the reservoir. The two objectives to be weighed are focusing on the safety of the reservoir itself and focusing on the safety of the flood control section downstream of the reservoir. If there are potential flood control risks in the flood control section downstream of the reservoir during the early risk discrimination in Step 1, and the reservoir itself is safe, then set the multi-objective decision-making weight W that prefers downstream flood control safety q ; if the downstream of the reservoir is safe, and there may be a risk of excessive water level in the reservoir itself, then set the multi-objective decision-making weight W that prefers the flood control safety of the reservoir z ; if there are potential flood control risks in both the downstream of the reservoir and the reservoir itself, then set the neutral multi-objective decision-making weight W e ; if there are no potential flood control risks in both the downstream of the reservoir and the reservoir itself, then initially set the neutral multi-objective decision-making weight W e .

[0083] In this embodiment, if there may be a risk of excessive reservoir water level while the downstream is safe, then determine the weight W z = 0.8:0.2; if the reservoir itself is safe while there may be a risk of flow exceeding the guarantee downstream, then determine the weight W q = 0.2:0.8; if there are no risks or both may have risks in the reservoir and the downstream section, the neutral weight W e = 0.5:0.5 or be determined according to the preference of the decision maker.

[0084] S33. Determine the flood control dispatching scheme for each reservoir:

[0085] For each reservoir in the reservoir group, after determining the multi-objective decision-making weight, use the augmented scale function decomposition method (ASF) to select the dispatching scheme that meets the decision maker's preference from the set of feasible dispatching schemes;

[0086] The formula of the ASF function is as follows:

[0087]

[0088] In the formula: f’ i (x) is the l-th feasible dispatching scheme; w is the decision-making weight; M is the number of feasible dispatching schemes; x is the feasible solution obtained by algorithm optimization; S t is the feasible region that meets the constraint conditions;

[0089] Select the smallest among the M AFS values as the dispatching scheme that best meets the decision maker's preference, and this scheme is used as the initial ideal flood control dispatching scheme for this reservoir.

[0090] Step 4, iterative optimization of flood control operation plan:

[0091] After determining the initial flood control operation plan, for each reservoir in the flood control target basin, input the current forecast inflow process to simulate the water level process during the reservoir operation period, compare it with the historical reservoir water level process under historical flood events, and calculate the differences in the highest water level, lowest water level, and end water level.

[0092] Use the Scheduling scheme iterative optimization algorithm (SSIOA) to adjust the initial flood control operation plan to make the highest water level, lowest water level, and end water level of the reservoir water level process under the action of the current forecast inflow process consistent with the historical process, and obtain the optimized flood control operation plan.

[0093] The specific implementation method of the SSIOA algorithm is as follows:

[0094] The input variables of the algorithm are the historical operation plan qh, the historical reservoir water level process Zh, the current calculated reservoir water level process Zc, the reservoir water level - storage capacity conversion equation HV(), the acceptable water level error dz, the scheduling unit time step ts, the reservoir water level calculation model Z = R(F, q), and the forecast inflow process F. The output variable is the optimized operation plan qo.

[0095] First, calculate the minimum water level difference dZmin, maximum water level difference dZmax, and end water level difference dZend between Zc and Zh as the initial optimization objectives; use HV() to calculate the target minimum storage capacity ObjVmin, maximum storage capacity ObjVmax, and end storage capacity ObjVend, initialize the optimized operation plan as qh, and initialize the optimized water level process as Zc.

[0096] Step 2: Execute loop iteration until dZmin, dZmax, and dZend are all less than the acceptable water level error dz. The loop iteration algorithm divides the water level process into three parts: the pre-drawdown stage before the flood peak arrives, the water level rising stage when the flood peak arrives, and the water level falling stage after the flood peak. First, handle the first part: Calculate the moment tmin when the lowest water level in Zc appears. Excluding the first moment and the tmin moment, all moments during this period are optimizable moments, and the number of optimizable time periods is Nmin = tmin - 2. The goal of the algorithm is to optimize the scheduling plan for Nmin scheduling periods to make up for dZmin. Calculate the current calculated lowest water level VminC. Assume the storage difference is dVmin = VminC – ObjVmin. Then the corrected discharge for each scheduling period is dqmin = (dVmin / Nmin) / (ts * 3600). Superimpose the corrected discharge on the initialized scheduling plan to obtain the updated scheduling plan. After finishing the calculation of the first stage, it is necessary to update the current calculated water level using the reservoir water level calculation model, and then enter the second stage of calculation. The second stage is similar to the first stage. First, calculate the storage VmaxC corresponding to the highest water level of the updated calculated water level process and calculate the water volume difference dVmax = VmaxC – ObjVmax. Calculate the moment tmax when the highest calculated water level appears. The allocable time period is the time period between tmin and tmax, and the number of time periods is Nmax = tmax – tmin – 1. Calculate the corrected discharge for each scheduling period dqmax = (dVmax / Nmax) / (ts * 3600). Finally, superimpose the corrected discharge on the optimized scheduling process to obtain the updated optimized scheduling plan, and update the current calculated water level using the reservoir water level calculation model. The third stage is from tmax to the end of the scheduling period, and the number of allocable time periods is Nend = N – tmax – 1, where N is the total number of scheduling periods. After calculating the water volume difference dVend = VendC – ObjVend in the same way as the previous steps, calculate the corrected discharge dqend = (dVend / Nend) / (ts * 3600) and superimpose it on the optimized scheduling process to obtain the updated scheduling plan. Finally, update the current calculated water level using the reservoir water level calculation model. If the termination requirement of the loop iteration is met, jump out of the loop and output the optimized scheduling plan qo; otherwise, loop for optimization again.

[0097] Step 5, Risk feedback and scheme iterative optimization of the reservoir group system:

[0098] S51, Establish and train a dynamic Bayesian network model for flood risk discrimination:

[0099] Taking all the reservoirs and key flood control sections in the flood control target basin as the key flood control nodes, a dynamic Bayesian network model is constructed, and the key flood control nodes are used as the nodes of the dynamic Bayesian network; taking the historical scenarios including historical flood events, feasible operation plans, reservoir operation water levels and downstream flood control section flow processes in step 3 as input variables, setting the safety boundaries of each flood control node, and training the dynamic Bayesian network model by the Monte Carlo simulation method to calculate the risk probabilities of each node in the historical flood scenarios.

[0100] S52, Real-time risk inference based on dynamic Bayesian network:

[0101] When a new flood event is about to occur, taking the current forecast flow process, the optimized flood control operation plan obtained in step 4, and the downstream section flow process as input, the trained dynamic Bayesian network model is used to infer the risk probabilities of each flood control node in the basin under the current forecast flood scenario.

[0102] S53, Risk feedback and iterative optimization of operation plan:

[0103] Set the safety threshold of acceptable risk. If the inference result of the dynamic Bayesian network model shows that the risk probability of a certain node exceeds the safety threshold and it is judged that there is a flood control risk at this node, then new reservoir water level and flow boundaries are formulated, and the iterative optimization algorithm is used to re-optimize the operation plans of each reservoir. Specifically:

[0104] If the flood control risk probability of a certain reservoir is relatively high, the operation water level of this reservoir needs to be reduced; the operation plan loop optimization algorithm SSIOA constructed in step 4 is used to reduce the target optimization water level and recalculate the operation plan;

[0105] If the peak flow of a certain section is too large, the outflow process of the reservoir related to this section needs to be optimized; the release optimization algorithm (ROA) is used to recalculate the operation plan to make the optimized plan meet the set flow boundary;

[0106] After the operation plan is optimized, it is substituted into the dynamic Bayesian network model again for risk inference, and the above steps are repeated until the potential risks of all nodes are within the acceptable range, thus completing the intelligent decision-making of flood control operation for the reservoir group system in the basin. Among them, SSIOA and ROA can be used alternately to obtain the results where both the reservoir water level and the operation rules meet the constraint boundary conditions.

[0107] The calculation process of the ROA algorithm is as follows:

[0108] The input variables of the algorithm are the number of reservoirs N hydraulically connected to the downstream section, the scheduling scheme matrix qMat = [q1, …, qN], the upper limit of the downstream section flow rate qm, the inflow process ql in the reach, and the Muskingum algorithm M(qi, pi), where qi is the discharge process of the i-th reservoir and pi is the parameter group of the algorithm; the matrix pMat = [p1, …, pN] composed of the Muskingum algorithm parameters from each reservoir to the downstream section. The output variables are the matrix qMat2 composed of the optimized scheduling scheme and the unit peak reduction dq.

[0109] Substitute the initial scheduling scheme of each reservoir into the Muskingum algorithm to calculate the flow process qc = M(qMat, pMat) + ql at the downstream section. When the peak flow rate is greater than the flow rate upper limit qm, perform iterative loop. Loop through each reservoir. Assume that the peak discharge of the current scheduling scheme is qmax and the peak occurrence time is ti. Initially set the target of reducing the peak discharge to qmax – dq. When the peak discharge of the scheduling scheme is greater than the target, perform the loop: Determine that all scheduling periods except the first and last periods and the ti period are allocable periods, and the number of allocable periods is n = length(qi) – 3. Allocate the reduced dq flow into each allocable period to ensure water balance, and the flow increment for each period is dqc = dq / n. After updating the scheduling scheme, detect whether other peak discharges exceeding qmax – dq are generated by the flow allocation. If so, continue the iteration; otherwise, end the scheme iteration for this reservoir.

[0110] After completing the iterative optimization of the schemes of N reservoirs, use the Muskingum algorithm to calculate the flow process at the downstream section of the updated scheduling scheme matrix, and calculate the peak flow rate after superimposing the inflow ql in the reach. If the peak flow rate is less than qm, end the iteration and output the updated scheduling scheme matrix qMat2; otherwise, continue the iterative loop.

[0111] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, those of ordinary skill in the art should understand that the technical solution of the present invention (such as the application of various formulas, the sequence of steps, etc.) can be modified or equivalently replaced without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An intelligent decision-making method for flood control dispatching of a river basin reservoir system, characterized in that: The method comprises the following steps: Step 1: Advance identification of flood control risks in the basin: First, determine the scope of the flood control target basin and identify the key flood control nodes within the flood control target basin; Then, by matching the relationship between the historical data of cumulative precipitation forecasts in the flood control target basin and the actual disaster situation of key flood control nodes, a historical flood risk event dataset was constructed; a deep residual network model was constructed, and the model was trained and verified on the flood risk event dataset; Input the real-time cumulative precipitation forecast data of the flood control target basin into the above trained and verified model to complete the prediction of potential flood control risks in the future; Step 2, identification of similar flood scenes: Collect the measured data of historical flood sites in the flood control target basin, establish a historical flood scenario database, and obtain the historical generalized flood through the mixed Gamma function fitting and normalization processing; Select Pearson correlation coefficient ρ, Euclidean distance ρ eu , dynamic time warping DTW and mean absolute error MAE as flood similarity evaluation indicators, calculate the current forecast flood and the current generalized flood; then calculate the similarity between the current generalized flood and the historical generalized flood, and match the historical flood event with the highest similarity to the current forecast flood; Step 3, preliminary selection of flood control scheduling scheme: A multi-objective optimization algorithm is used to solve offline the feasible dispatching scheme set of each reservoir in the reservoir group in each historical flood event, and the scheme is stored in the historical flood scenario database to obtain the dispatching plan library; When a new flood event is about to occur, the method in step 1 is first used to pre-judge the flood control risk of the basin, and then the historical flood event with the highest similarity to the current flood forecast is matched according to the method in step 2, and the feasible scheduling plan set is matched at the same time; the subjective weight of the decision is determined according to the risk prediction of each reservoir itself or the downstream of the reservoir, and the augmented scaling function decomposition method is used to select the flood control scheduling plan that best meets the decision maker's preferences from the scheduling plan library, complete the preliminary selection of the flood control scheduling plan, and obtain the initial flood control scheduling plan; Step 4: Iterative optimization of flood control scheduling scheme: After the initial flood control scheduling scheme is determined, for each reservoir in the flood control target basin, the current forecast inflow process is input to simulate the water level process during the reservoir scheduling period, and compared with the historical reservoir water level process under historical flood events to find the difference between the highest water level, the lowest water level and the end-of-period water level; The initial flood control scheduling scheme is adjusted using the scheduling scheme cyclic optimization algorithm to make the highest water level, the lowest water level and the final water level of the reservoir water level process under the current forecast inflow process consistent with the historical process, and the optimized flood control scheduling scheme is obtained; Step 5: Risk feedback of reservoir group system and iterative optimization of solutions: Taking all reservoirs and key flood control sections in the flood control target basin as the key flood control nodes, a dynamic Bayesian network model is constructed, and the key flood control nodes are used as nodes of the dynamic Bayesian network; taking the historical scenarios including historical flood events, feasible scheduling schemes, reservoir operating water levels and downstream flood control section flow processes in step 3 as input variables, the Monte Carlo simulation method is used to train the dynamic Bayesian network model; In real-time flood events, the current forecast flow process, the optimized flood control scheduling scheme obtained in step 4, and the downstream section flow process are used as inputs, and the dynamic Bayesian network model is used to infer the risk probability of each flood control node in the basin; A safety threshold for acceptable risk is set. If the reasoning result of the dynamic Bayesian network model shows that the risk probability of a node exceeds the safety threshold, it is judged that the node has a flood control risk. Then a new reservoir water level and flow boundary is formulated, and the scheduling scheme of each reservoir is re-optimized using the scheduling scheme cyclic optimization algorithm. After optimization, the dynamic Bayesian network model is substituted again for risk reasoning, and the above steps are repeated until the potential risks of all nodes are within an acceptable range, thereby completing the intelligent decision-making of flood control scheduling of the basin reservoir group system.

2. The intelligent decision-making method for flood control dispatching of a river basin reservoir system according to claim 1 is characterized in that: Step 1 The specific steps of advance identification of flood control risks in the basin include: S11, Analyze the historical grid data of basin cumulative precipitation forecast: Determine the scope of the flood control target basin, obtain and analyze the historical data of 3-hour, 6-hour and 12-hour cumulative precipitation forecasts in the basin, and process them into a 2D raster image in .tif format, i.e., a historical cumulative precipitation forecast data map; S12, determine the key flood control nodes in the basin and determine the risk quantification standards: Take the locations of key flood control sections and water projects within the flood control target basin as key flood control nodes, and determine the safety standards for each key flood control node; S13, cumulative precipitation historical forecast data map matched to each risk category according to time: First, the measured data of historical flood sites in the flood control target basin are collected, and the risk scenarios are classified according to whether the key flood control nodes have experienced risks in historical flood events: if node i has experienced an excessive flood risk in historical flood events, it is defined as category i-0; if both node i and node j have experienced an excessive flood risk, it is defined as category ij-0; if no node has an excessive flood risk, it is defined as category none; The forecast data graph processed by S11 is matched with historical flood events according to the occurrence time and placed into each category to construct a historical flood risk dataset; S14, training a deep residual network model using a historical flood risk dataset: Construct a deep residual network model Resnet-50 and split the historical flood risk dataset into a training set and a validation set; train the Resnet-50 network on the training set and optimize the hyperparameters to ensure that the prediction accuracy of the model on the validation set reaches more than 95%; S15, input real-time forecast data into the model and predict future flood control risks: The real-time cumulative precipitation forecast data of the flood control target basin is obtained, processed into a two-dimensional raster image in .tif format according to S11, and then input into the trained deep residual network model Resnet-50 to predict the potential flood control risks of key flood control nodes in the basin.

3. The intelligent decision-making method for flood control dispatching of a river basin reservoir group system according to claim 2 is characterized in that: In step 1 S21, the key flood control nodes include densely populated cities built on water and small and medium-sized reservoirs prone to flood risks; the safety standards of the key flood control nodes include the safe flow of the river section, the safe water level, and the safe water level of the reservoir.

4. The intelligent decision-making method for flood control dispatching of a river basin reservoir system according to claim 1 is characterized in that: Step 2: The specific steps of identifying similar flood scenes include: S21, establish a historical flood scenario database: Determine a fixed forecast period as the flood duration, match the historical inflow process of each reservoir and the historical interval inflow process from the reservoir to the downstream section in time, and store them together to establish a historical flood scenario database; S22, Generalization of flood events: First, the flood events are normalized. The calculation formula of the normalization method is: Where: q(t) is the normalized flow value; Q(t) is the original flow value, m 3 / s;Q max and Q min are the maximum and minimum values ​​of the original flow series, m 3 / s; t' is the normalized time; t is the original time, h; LT is the forecast period, h; Then, according to the differences in characteristics between the flood stage and the receding stage, the flood process was fitted based on the mixed Gamma function to obtain the generalized flood process. The formula for the mixed Gamma function is: Where: k1 and k2 are normalization constants to ensure the continuity of the two-stage function process; t' rise and t' fall are the time scales of the flood and receding stages respectively; t' peak is the time when the flood peak occurs; α1 and α2 are the shape parameters of the Gamma function in the flood stage and the receding stage respectively, Γ() is the Gamma function; e is a natural constant; y() is the generalized flood process line; The parameters in the mixed Gamma function are calibrated using the NSGA-III algorithm, and three objective functions are set: F1=min{q fit -q obs } <h2 style=";text-align:left;direction:ltr">F2 = min {w<h2 style=";text-align:left;direction:ltr"> fit <h2 style=";text-align:left;direction:ltr"> -w<h2 style=";text-align:left;direction:ltr"> obs <h2 style=";text-align:left;direction:ltr">} F3=max{NSE(fit,obs)} Where: F1, F2 and F3 are three calibration targets; q fit and q obs are the normalized generalized flood peak size and the normalized measured flood peak size respectively; w fit and w obs are the normalized generalized flood volume and the normalized measured flood volume respectively; fit and obs are the normalized generalized process and the normalized measured process respectively; NSE() is the Nash-Sutcliffe coefficient; The generalized flood process and the original flood process are stored together in the historical flood scenario database; S23, construct similarity evaluation index system: Select Pearson correlation coefficient ρ, Euclidean distance ρ eu , dynamic time warping DTW and mean absolute error MAE are used as evaluation indicators, and the calculation formulas are as follows: Where: f and his are the generalized process of the forecast flow process and the generalized process of the historical flow process, respectively; cov() is the covariance, σ f and σ his are the standard deviation of the forecast flow process and the standard deviation of the historical flow process respectively; N is the total number of time periods; q f (t') and q his (t') are the generalized flow values ​​of the forecast process and the historical process at time t' respectively; DTW() is the dynamic time warping algorithm, K is the total number of paths, w k is the kth subpath; S24, Identification of Similar Flood Scenes: Based on the forecast precipitation process, the forecast flow process is calculated using the hydrological model; the current generalized flood is calculated using the S22 method, and then the S23 index system is used to calculate the similarity index value between the current generalized flood and each historical generalized flood, and the calculated index value is normalized; the Pearson correlation coefficient is a positive index, and the Euclidean distance, dynamic time bending and mean absolute error are negative indicators; the positive index remains unchanged, and the negative index is reversed; each index value is given a subjective weight W T , get the similarity score; Conduct similarity retrieval of flood events, that is, search the historical flood generalization process that is most similar to the current forecast flood generalization process in the historical flood scenario database according to the similarity score. This is divided into the following three steps: (a) Select the 20 historical flood events with the highest similarity scores to form flood event set A1; (b) Eliminate flood events with relative errors of flood peaks greater than 30% from A1 to obtain flood event set A2; (c) Eliminate flood events with relative errors of flood volume greater than 30% in A2 to obtain flood event set A3; Finally, the historical generalized flood with the highest similarity score in A3 is selected, and the historical flood event that is most similar to the current forecast flood event is obtained after the process is denormalized.

5. The intelligent decision-making method for flood control dispatching of a river basin reservoir system according to claim 1 is characterized in that: Step 3 The specific steps for the preliminary selection of flood control scheduling schemes include: S31, offline calculation of multi-objective optimization flood control scheduling schemes for each reservoir under historical flood event scenarios: For each reservoir in the reservoir group, the multi-objective optimization dispatching scheme in each historical flood event is calculated to obtain a set of feasible dispatching schemes; the feasible dispatching schemes are matched with flood events and stored in the historical flood scenario database to form a dispatching plan library; S32, determine the decision preference weights of each reservoir: According to the method in step 1, the flood control risk of the basin is judged in advance, and then according to the method in step 2, the historical flood events with the highest similarity to the current flood forecast are matched, and the feasible scheduling scheme set is matched at the same time; If, in the risk advance judgment of step 1, the flood control section downstream of the reservoir has potential flood control risks, and the reservoir itself is safe, then the multi-objective decision weight W that prefers downstream flood control safety is set q If the downstream of the reservoir is safe, but the reservoir itself may face the risk of excessively high water levels, then the multi-objective decision weight W that prefers reservoir flood control safety is set. z ; If both the downstream of the reservoir and the reservoir itself have potential flood control risks, then a neutral multi-objective decision weight W is set e ; If there is no potential flood control risk in the downstream of the reservoir and the reservoir itself, a neutral multi-objective decision weight W is initially set e ; S33, determine the flood control dispatching plan for each reservoir: For each reservoir in the reservoir group, after determining the multi-objective decision weights, the augmented scaling function decomposition method is used to select the scheduling scheme that meets the decision maker's preference from the feasible scheduling scheme set according to the weights; The formula of the augmented scaling function decomposition function is as follows: Where: f' i (x) is the lth feasible scheduling scheme; w is the decision weight; M is the number of feasible scheduling schemes; x is the feasible solution obtained by algorithm optimization; S t is the feasible domain that meets the constraints; The smallest of the M AFS values ​​is selected as the scheduling scheme that best meets the decision maker's preference, and this scheme is used as the initial ideal flood control scheduling scheme for the reservoir.

6. The intelligent decision-making method for flood control dispatching of a river basin reservoir system according to claim 1 is characterized in that: Step 5: The specific steps of risk feedback of reservoir group system and iterative optimization of solutions include: S51, establish and train a dynamic Bayesian network model for flood risk identification: Taking all reservoirs and key flood control sections in the flood control target basin as the key flood control nodes, a dynamic Bayesian network model is constructed, and the key flood control nodes are used as nodes of the dynamic Bayesian network; taking the historical scenarios including historical flood events, feasible scheduling schemes, reservoir operating water levels and downstream flood control section flow processes in step 3 as input variables, the safety boundaries of each flood control node are set, and the dynamic Bayesian network model is trained using the Monte Carlo simulation method to calculate the probability of risk of each node under the historical flood scenario; S52, Real-time risk reasoning based on dynamic Bayesian networks: When a new flood event is about to occur, the current forecast flow process, the optimized flood control scheduling plan obtained in step 4, and the downstream section flow process are used as inputs, and the trained dynamic Bayesian network model is used to infer the risk probability of each flood control node in the basin under the current forecast flood scenario; S53, risk feedback and scheduling scheme iterative optimization: Set a safety threshold for acceptable risk. If the reasoning result of the dynamic Bayesian network model shows that the risk probability of a node exceeds the safety threshold, it is judged that the node has a flood control risk. Then, a new reservoir water level and flow boundary is formulated, and the iterative optimization algorithm is used to re-optimize the scheduling plan of each reservoir. Specifically: If the flood risk probability of a reservoir is high, the operating water level of the reservoir needs to be lowered; the scheduling scheme loop optimization algorithm constructed in step 4 is used to lower the target optimized water level and recalculate the scheduling scheme; If the peak flow of a certain section is too large, it is necessary to optimize the outflow process of the reservoir related to the section; use the outflow optimization algorithm to recalculate the scheduling plan so that the optimized plan meets the set flow boundary; After the scheduling plan is optimized, it is substituted into the dynamic Bayesian network model again for risk reasoning, and the above steps are repeated until the potential risks of all nodes are within an acceptable range, thereby completing the intelligent decision-making of flood control scheduling in the basin reservoir group system.

Images