[0066] The reservoir group scheduling decision behavior excavation method and the reservoir schedule automatic control device according to the present invention will be described in detail below with reference to the accompanying drawings.
[0067]
[0068] like figure 1 As shown, the reservoir group scheduling decision behavior mining method provided in this embodiment includes the following steps:
[0069] Step 1, determine the research scenario, collect the basic information of the reservoir group system, historical scheduling data and the library site meteorological data, and determine the influence factor and decision variables of the reservoir dispatching behavior.
[0070] Step 2: Determining a depth learning algorithm for excavating the scheduling decision making behavior of the reservoir, such as a long short-term memory network (LSTM) model.
[0071] Step 3: Review the accuracy of the scheduling data of the reservoir group, that is, the flow rate of the various reservoirs, the warehousing flow and the water level in the water level in the water volume equation. If the data is accurate, the time series data of the spring sound factor in step 1 is normalized, and the training set and test set is divided according to the length of time sequence data, and the time series of the first 60% length is selected as the training set.
[0072] Step 4, constructing the basic principles of coupling reservoir and the reservoir group scheduling decision behavior excavation model of depth learning technology, and simultaneously use the design flooding process and the historical actual operation data as a model training sample, using the design flood scheduling model to broaden deep learning The model of the model, promotes the depth learning model to absorb the basic principle of the reservoir scheduling; in addition, the upper and lower boundary constraints, water volume balance constraints, monotonic constraints of the reservoir scheduling decision, and cause the scheduling behavior of the excavation of the reservoir scheduling basics. principle.
[0073] Based on the scheduling model of design flooding model, the training samples of deep learning model are widened. In addition to historical scheduling information and hydrological meteorological data, the depth learning model can learn the basic principle of reservoir scheduling by absorbing the reservoir scheduling physical model, thereby the basic principle of coupling reservoir scheduling. First, the design flood information of each reservoir of the reservoir is collected in the frequency of 20% to 0.01%. Then, comparing the flood control capacity of the various reservoirs in the reservoir, considering the reservoir mainly assumed flood control task in the cascade reservoir, with the design floods in the reservoir, in the basis of the warehouse, the warehouse flow, the warehouse flow, the warehouse, which is treated with the watershed. Further, according to the corresponding warehousing runoff, each of the reservoirs will be carried out, and the regular reservoir scheduling process is obtained. Finally, the information such as the warehouse flow, the water level, the water level, and the reservoir force, and provides dispatching cases that contain flood scheduling laws for deep learning model training.
[0074] In addition, the upper and lower boundary constraints, water volume balance constraints, monotonic constraints, monotonic constraints, monotonic constraints in the depth learning model, and the three constraints are coupled to the depth learning model as follows:
[0075] The upper and lower boundary constraints of the reservoir scheduling decision are the upper boundary and lower boundary constraints of the reservoir group. If the simulation value of the reservoir group outbound storage is not in line with the boundary constraint, the corresponding base value Loss is required. boundary , Specifically, the following formula is calculated:
[0076] have:
[0077] which is,
[0078] Where N is the number of reservoirs in the reservoir; n t For the time series length of the test period; Simulation of the flow rate of the warehouse flow for the J20 of the reservoir of the reservoir, M 3 / S; Q l The lower limit of the flow of the reservoir, M 3 / S; Q u The upper limit of the flow of the reservoir, M 3 / s.
[0079] The water volume balance constraint of the reservoir scheduling decision is calculated by calculating the simulation value of various reservoirs in the reservoir, and the non-equilibrium value of the reservoir reservoir changes. balance To achieve, specifically, the following formula is calculated:
[0080]
[0081] In the formula, ΔW j Changes to the water volume of the reservoir in the reservoir, M 3 Q t，j The warehousing flow of the Journal of the reservoir of the reservoir is calculated from the real value and interval of the outbound flow of the upstream reservoir, M 3 / S; Simulation of the flow rate of the warehouse flow for the J20 of the reservoir of the reservoir, M 3 / s.
[0082] The monotonic constraint of the reservoir scheduling decision is the other input sequence of the model, and when the storage flow sequence increases a certain value, the reservoir outbrace flow simulation result is not less than the simulation result of the original input sequence. If the analog value of the reservoir group outbound flow is not in line with the monotonic constraint, the non-monodulated value LOSS is calculated. monotony , Specifically, the following formula is calculated:
[0083]
[0084] In the form of For the configured assumption, other input conditions are unchanged, and the flow rate flow sequence simulation value is increased when the flow rate is increased. 3 / S; Original flow simulation value for the original input sequence, M 3 / s. RELU is a non-linear activation function, RELU (z) = z + = Max (0, z).
[0085] After calculating the projection value of the dispatch decision and the non-single-quoted value, the weighted summation is incorporated into the loss function of the long short-term memory network model, which is used for the reverse propagation of the loss function, which promotes the network update parameters and ultimately converges.
[0086] Step 5, based on the super parameters of the training set, including the maximum iterative number, the number of hidden layers, the number, the batch value, etc., based on the network parameters of the model loss function reverse propagation update model, and finally established the reservoir dispatch The influencing factor of behavior and the mapping relationship of decision variables, and realize the excavation of reservoir group scheduling decision behavior. Map of influence factors and decision variables for the scheduling behavior of the reservoir group:
[0087]
[0088] In the formula, F is the mapping relationship between the influence factor and the decision variable, and n is the number of reservoirs in the reservoir.
[0089] In this embodiment, the overall evaluation index index of the stream simulation accuracy of the reservoir group (Nash-Sutcliffe Efficiency Coefficient, referred to as: NSE) is calculated according to the following formula:
[0090]
[0091] Q obs (i) is the actual survey of the model test period, Q sim (i) is the simulated outlet flow of the model test period, Is the average value of the model test period, N t The length of the time series for the test period.
[0092] As a specific embodiment, first, the ultra-parameters of the training collector are finalized, the final determined training iteration is 200, the number of hidden layers is 2, the hidden layer node is 32, the lag time is 20, the learning rate For 0.005, the number of samples is 365.
[0093] In a specific embodiment, the reservoir dispatching decision-making behaviors using the coupling reservoir scheduling basic principles and deep learning techniques, excavation of Qingjiang River Cladding, Zhiri Reservoir 2009 ~ 2019 Day scale outbound traffic Decision, the acquisition of historical scheduling data contains time, water cloth 垭 垭 流 流, hydrangea - sessage flow rate, water cloth and segregation of the rock out of the warehousing flow, water cloth and across the river reservoir upstream water level . After school, the scheduling data is in line with the water balance equation.
[0094] Two sets of model inputs are set by two application scenarios, and the model output is a water cloth and the current time of the crosstalk, find the mapping relationship between the model input and output, thus excavating the scheduling decision making behavior of the reservoir group.
[0095] Considering the refactoring of the reservoir historical export decision, since the reservoir has only timed, store traffic and interval flow data, the model input can only consider the above variables, but the scheduling decision needs to consider the water level of the reservoir Changes, therefore, calculate the average daily average water level of water cloth and septum rock, as a reference for the water level on the water level, thereby determining the model input. Therefore, the first set of models are input to timing, the water cloth in the current period of the reservoir, the flow rate, the interval flow, water cloth and the river rock in the current period of the water, the average water level in the river rock.
[0096] Considering the second application scenario is a forecast for the future flow of the reservoir group, therefore, the current available information is input, so the second set of models are input to timing, the water cloth in the current period is in the library traffic and water cloth 垭 - River rock flow, water cloth and sections of the reservoir in the reservoir of the reservoir, water cloth, and a time period of the next time in the river.
[0097] from figure 2 It can be seen that under the input of two application scenarios, the long-term memory network model of the coupling reservoir scheduling basic principle is a long short-term memory network model simulation result (LSTM)'s overall accuracy. improve. Under the first set of model inputs, the NSE of the watery reservoir outbound flow simulation is raised from 0.5846 to 0.7028, and NSE of the river rock reservoir outflow flow simulation is raised from 0.5564 to 0.7240; under the second set of model inputs, water cloth The NSE of the warehousing flow simulation was raised from 0.8189 to 0.8701, and NSE of the flow simulation of the river rock reservoir was raised from 0.8267 to 0.8665. When the input information is small, the effect of the model coupling reservoir scheduling principle is more pronounced.
[0098] As can be seen from Figures 3 (a) and 3 (b), the analog traffic of the long short memory network model (LSTM) has an underestimated phenomenon of high value outflow traffic, and the long-term memory network model of the coupling reservoir scheduling basic principle (PHY -Lstm can estimate the high value outlet flow, so the PHY-LSTM model enhances a cognition of scheduling behavior under extreme hydrological conditions.
[0099] As can be seen from Fig. 4 (a) and 4 (b), the analog traffic of long short-term memory network model (LSTM) still has a low-ranking high value outbound traffic (most high value outflow traffic is below Y = X) . Under the second set of model inputs, the simulation results of the water cloth reservoir LSTM model have no negative traffic, and the simulation results of the PHY-LSTM model do not have a negative traffic. Therefore, the long short-term memory network model (PHY-LSTM) of the coupling reservoir scheduling basic principles can simulate the flow decisions of the reservoir group outlet, and then extract the reservoir group scheduling behavior.
[0100] Depend on Figure 5 (a) ~ 5 (d) It can be seen that under the application scenario input by the first set of models, compared with the reservoir routine scheduling simulation results, the long-term memory network model (LSTM) reservoir outbound flow reconstruction results cannot adapt to the seasonal changes in the storage traffic of the reservoir, presenting itself Related features; long-term memory network model (PHY-LSTM) of the coupling reservoir scheduling basic principles, the resource reconstruction results can adapt to the seasonal changes in the warehousing flow of the reservoir, and better have learned the annual adjustment and multi-year regulation reservoir. Advantages.
[0101] As can be seen from Figures 6 (a) and 6 (b), under the application scenario input by the second set of models, the long short-term memory network model of the coupling reservoir scheduling base principle is analog results (PHY-LSTM). The simulation accuracy of a longer short memory network model simulation result (LSTM) is used separately, so the long-term memory network model (PHY-LSTM) of the basic principle of the coupling reservoir scheduling is more than the input data and output data. steady.
[0102] The basic principle of the coupling reservoir provided by the embodiment of the present invention and the reservoir group scheduling decision behavior mining method of deep learning techniques will be based on the design flood schedule process and historical actual operation data as a model training sample, and promote depth learning model to absorb the reservoir scheduling basics. Principle, enhance the depth learning model to scheduling behavior under extreme hydrological conditions; in addition, the upper and lower boundary constraints, water volume balance constraints, monotonic constraints are coupled in the deep learning model, and the scheduling behavior of excavation is more in line with the reservoir Scheduling basic principle. The present invention can be widely used in the scheduling behavior excavation of complex reservoir group systems, and the experience of scheduling personnel can be used to provide technical support for clutch reservoirs.
[0103] Further, the present embodiment also provides a reservoir group scheduling automatic control device capable of automatically implementing the above method, the device comprising a data acquisition unit, an algorithm determining unit, a calibration unit, a coupling model member, a mapping relationship establishment, a reservoir scheduling unit, input The display unit, the control unit.
[0104] The data acquisition department can determine the research scenario, obtain the basic information of the reservoir group system, the historical scheduling data and the reservoir site meteorological data, and determine the impact factor and decision variables of the reservoir restriction behavior; the factor of the reservoir dispatching behavior includes the following variables Or part of the variable: timing T, current period faucet water storage traffic Q in，L T , Current period reservoir group I reservoir to the JD library to enter the stream Q in，i，j T , Upstream water level of the reservoir in the reservoir in the previous time u，i T , Downstream water level in the reservoir of the reservoir in the last time d，i T , The outbound flow rate Q in the reservoir of the reservoir in the last time out，i T-1 , The previous period of the reservoir group I reservoir's output N out，i T-1 , Current time period reservoir group I reservoir rainfall And the evaporation amount E of the current period reservoir group I reservoir i T The decision variable of the reservoir group scheduling behavior is the current period of the reservoir in the current period. out，i T.
[0105] The Algorithm Determination Department determines the depth learning algorithm for excavating the scheduling decision making behavior of the reservoir group.
[0106] The calibration department is used to review the accuracy of the scheduling data of the reservoir group: check whether the flow rate of each reservoir, the warehouse flow rate and the water level in the water volume balance equation, if the data is accurate, the influence factor determined in step 1 is The time series data of the decision variable is normalized, and the training set and test set is divided according to the length of time sequence data.
[0107] Coupled model components for constructing the basic principles of coupling reservoirs and reservoir group scheduling decision behavior excavation models of deep learning models.
[0108]The mapping relationship establishment unit is based on the super parameters of the training set rate, including the maximum number of iterations, implicit layer nodes, learning rate, bulk value, and the network parameters of the model loss function reverse propagation model, according to the model in the test set. The simulation accuracy determines the optimal super parameters of the model, ultimately establish the mapping relationship between the influence factors and decision variables of the following reservoir group, and realize the mining of reservoir dispatch decision-making behaviors;
[0109]
[0110] In the formula, F is the mapping relationship of the factor and the decision variable, and n is the number of reservoirs included in the reservoir group.
[0111] The Reservoir Scherage determines the corresponding dispatching measures based on the mapping relationship determined by the mapping relationship establishment, and controls the reservoirs in the reservoir in accordance with the dispatching measures.
[0112] The input display unit communicates with the control unit to allow the user to enter an operation instruction and perform the corresponding display. For example, the input display unit can display the mapping relationship established by the mapping relationship establishment unit according to the operation instruction, and can display the schedule measures determined by the reservoir schedule and the operation of each reservoir according to the operation instruction. Scheduling measures include controlling the amount of flow, traffic, and water storage amount of the reservoir, and the amount of power generation.
[0113] The control unit is connected to the data acquisition unit, the algorithm determining unit, the calibration unit, the coupling model member, the mapping relationship, and the reservoir scheduling unit, and the input display unit communicates, and controls their operation.
[0114] The above embodiments are merely illustrative of the technical solutions of the present invention. The reservoir group scheduling decision behavior excavation method and the reservoir schedule automatic control device according to the present invention are not limited to the contents described in the above embodiments, but are scope as defined in the claims. Any modified or supplemental or equivalent replacement of the present invention, in the art, in the scope of the claims.
