Complex city river network hydrological data prediction method, optimal scheduling method and system

Abstract

The invention discloses a complex city river network hydrological data prediction method, optimization scheduling method and system, and the method comprises the steps: firstly constructing a space-time diagram neural network model, which comprises the steps: constructing a three-dimensional space-time diagram matrix reflecting the dynamic state of a river network and an adjacent matrix capable of aggregating the space-time information at the same time, and forming a complete space-time diagram data sample; extracting space-time fusion features by adopting a space-domain graph convolution method, and introducing an attention mechanism to calculate the correlation degree between nodes; and then training the space-time diagram neural network model, and carrying out multiple iterations to fit the model and achieve the optimum. And further according to predicted hydrological data obtained by the time-space diagram neural network model, based on the determined scheduling target and constraint conditions, constructing a river network optimization scheduling model, and solving the river network optimization scheduling model by adopting an NSGA-III algorithm to realize optimization scheduling of urban river network water environment improvement. The method has the advantages of being high in scheduling optimization quality and high in timeliness, and construction of a complex river network hydrological hydrodynamic physical model is avoided.

Description

Technical Field

[0001] The present invention relates to water resources optimization and scheduling, and in particular to a method for predicting hydrological data of a complex urban river network, an optimization and scheduling method and a system. Background Art

[0002] At present, the mainstream multi-objective scheduling methods of urban river networks include mathematical modeling and data-driven methods.

[0003] The mathematical modeling method models the urban river network, constructs a hydrological and hydrodynamic physical model of the river network, and finally obtains a hydraulic-based urban river network prediction model through continuous parameter calibration and model adjustment, which is used to guide the optimization and scheduling of the urban river network. Existing problems: The mathematical modeling method simulates the river network runoff according to the river hydrodynamics. The model construction process is complex and the calculation time is long. The obtained scheduling plan lacks timeliness and cannot meet the real-time optimization and scheduling needs.

[0004] The data-driven method is based on deep learning algorithms, such as genetic algorithms, artificial neural networks, particle swarm algorithms, ant colony algorithms, etc., which optimize the values ​​of decision variables to make the objective function optimal under given constraints. Existing problems: The data-driven method fails to integrate time and space dimensions and fully consider the spatiotemporal complexity of river network optimization scheduling. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method for predicting hydrological data of a complex urban river network, an optimization scheduling method and a system to solve one or more problems existing in the prior art.

[0006] Technical solution: A method for predicting hydrological data of a complex urban river network of the present invention comprises:

[0007] (1) Collect and preprocess the hydrological data of urban river networks to form a hydrological dataset;

[0008] (2) Establish a spatiotemporal graph neural network model;

[0009] (201) Taking hydrological stations and sluice pumps as nodes, and the water level, flow rate, and velocity data of the nodes as the characteristic values ​​of the nodes, the characteristic matrices of all nodes within a period of time are combined together to form a three-dimensional space-time graph matrix N;

[0010] (202) constructing a spatiotemporal adjacency matrix that can simultaneously aggregate spatiotemporal information on a node, combining the spatiotemporal adjacency matrix with the three-dimensional spatiotemporal graph matrix, and connecting each node through the spatiotemporal adjacency matrix to form a complete spatiotemporal graph data sample;

[0011] (203) The spatial domain graph convolution method is used to extract spatiotemporal fusion features from the spatiotemporal graph data samples, an attention mechanism is introduced to calculate the correlation between nodes, and a gating mechanism in the gated convolution is introduced, and a residual connection connected to the aggregation layer is introduced in each layer; the spatiotemporal fusion features are subjected to global maximum pooling to form the final feature vector and input into the fully connected layer, which uses the softmax activation function to output the classification results, namely the river network hydrological prediction data;

[0012] (3) The hydrological data set is used to train and optimize the spatiotemporal graph neural network model, and the trained spatiotemporal graph neural network model is used to predict hydrological data.

[0013] Further, step (1) comprises:

[0014] (101) Obtaining urban river network hydrological data

[0015] Obtain hydrological data from existing hydrological stations and sluice pumps; for key river network nodes where no water conservancy buildings or hydrological stations have been built, observe hydrological data through hydrological measurement equipment;

[0016] (102) Preprocess the acquired urban river network hydrological data to form a hydrological data set

[0017] Nodes were designed at key locations of the river network, the collected data were counted, and missing values ​​were filled in. The data were normalized using the deviation standardization method, and the conversion formula was:

[0018]

[0019] Among them, X * is the normalized data, which ranges from [0,1]; X is the original data; X max is the maximum value of the original data; X min is the minimum value of the original data.

[0020] Further, step (202) includes: generating a spatial adjacency matrix A of size N×N according to the real topological structure of the river network S , then construct the spatiotemporal adjacency matrix A ST , the spatiotemporal adjacency matrix A ST It consists of two sub-matrices, including the spatial adjacency matrix A S and the time adjacency matrix A T ; When considering processing s time steps in a graph convolution, the spatiotemporal adjacency matrix A of size sN×sN is obtained ST , A S Located on the diagonal, A T The nodes are located on both sides of the diagonal line.

[0021] Further, step (203) includes: using a spatial domain graph convolution method to perform graph convolution, through the feature matrix H l Matrix multiplication with the incidence matrix A aggregates information:

[0022] H l+1 =σ(AH l W l +b)

[0023] Among them, σ is the activation function; A is the adjacency matrix of the corresponding space-time graph; H l is the feature matrix of the lth layer; W l is the attention matrix of the lth layer; b is an additional bias used to fine-tune the results;

[0024] The attention mechanism is introduced to calculate the correlation between nodes. The attention matrix W is involved in the calculation as shown in the following formula:

[0025] W=Oσ(Y l-1 Z1Z2(Z3Y l-1 ) T +b)

[0026] Among them, O is the mapping function, Y l-1 is the output of the l-1th space-time block and the input of the lth space-time block; Z1, Z2, and Z3 are all learnable parameter matrices;

[0027] Multiple graph convolution modules are connected to form a spatiotemporal graph convolution combination module to expand the spatial receptive field. In the graph convolution process, a gated linear activation function is used as the activation function σ. The graph convolution process combined with the gating mechanism is described as:

[0028] H l+1 =(AH l W1+b1)⊙sigmoid(AH l W2+b2)

[0029] Where W1 and W2 are attention matrices; b1 and b2 are additional biases used to fine-tune the results; ⊙ represents the XOR operation; sigmoid(·) is a soft saturation activation function;

[0030] In addition to the gating mechanism, a residual connection connected to the aggregation layer is introduced in each layer, which is specifically described as:

[0031] H l+1 =F(H l ,A)+H l

[0032] Among them, F(H l ,A) indicates that H l The output feature matrix calculated by and A; H l+1is the result after addition.

[0033] Furthermore, step (3) includes: using mean square error as the loss function, calculating the gradient back-propagation to update the weights of each layer after calculating the loss, fitting the model and achieving the optimum through multiple iterations of gradient descent, using the Adam algorithm to update the weights according to these gradients, and using regularization methods to prevent overfitting.

[0034] The present invention provides a complex urban river network optimization scheduling method, comprising: constructing a river network scheduling optimization model based on the determined objective function and constraints according to the hydrological prediction data obtained by the above-mentioned complex urban river network hydrological data prediction method; and solving the river network scheduling optimization model using the NSGA-III algorithm to obtain a river network scheduling plan.

[0035] Furthermore, the objective function includes ecological flow rate and ecological water level:

[0036] (a) Ecological flow rate

[0037] v≥max(v1,v2)

[0038] Among them, v is the ecological flow rate;

[0039] v1 is the velocity target to ensure the flow of the river, and is calculated using the RVA method using the following formula:

[0040]

[0041] in, is the multi-year average monthly value of flow velocity, v 上限 and v 下限 The values ​​corresponding to 75% and 25% of the probability of occurrence of hydrological evaluation indicators in long-sequence years are selected respectively;

[0042] v2 is the flow velocity required by the living environment of urban river fish and the flow velocity when algae growth is inhibited;

[0043] When the ecological flow rate guarantee rate of a river reaches 50%, it is considered to meet the ecological flow rate requirements. The ecological flow rate guarantee rate of a river refers to the proportion of the duration greater than the ecological flow rate to the total duration;

[0044] (b) Ecological water level

[0045] H4>H≥max(H1,H2,H3)

[0046] Among them, H1 is the ecological water level required to maintain river connectivity, which is also calculated using the RVA method; H2 is the ecological water depth to maintain fish habitats, which needs to consider the survival depth required by fish at different growth stages, and the water depth required for fish to turn vertically; H3 is the ecological water level required to maintain good landscape entertainment factors, and the calculation formula is:

[0047] H3=max(H 通航 ,H 灌溉 ,H 亲水 )

[0048] Among them, H 通航 The lowest river water level required for navigation; H 亲水 The water level of the river is suitable for the water-friendly leisure activity platform; H 灌溉 Minimum water level required for irrigation on both sides of the river:

[0049] H 灌溉 =min(H 灌溉1 ,H 灌溉2 ,…,H 灌溉j )

[0050] Among them, H 灌溉j is the designed minimum water intake level of the jth water intake structure;

[0051] H4 is the river water level that should not be exceeded before the flood season for flood control and municipal drainage. The calculation formula is:

[0052] H4=min(H 防洪 ,H 市政 )

[0053] Among them, H 防洪 The flood control level stipulated by the water management department; 市政 To ensure stable water level of municipal drainage system:

[0054] H 市政 =min(H 排1 ,H 排2 ,…,H 排m )

[0055] Among them, H 排m is the water level elevation of the mth drainage outlet;

[0056] When the ecological water level protection rate of the river reaches 50%, it is deemed to meet the ecological water level requirements. At the same time, the water level limit for flood control and drainage shall not be exceeded at any time.

[0057] Furthermore, the constraints include water balance constraints of sluice pumps, flow and velocity constraints of various levels of river channels, and engineering operation constraints:

[0058] (a) Water balance constraints of sluice pumps

[0059] Q b,in,(k,t) =Q b,out,(k,t)

[0060] Among them, Q b,in,(k,t) represents the water diversion flow of the sluice pump during period t, Q b,out,(k,t) Indicates the water flow rate of the sluice pump during period t, m 3 / s;

[0061] (b) Constraints on flow and velocity at all levels of rivers

[0062] Q c,up,(t) =Q c,down,(t) +L c,(t)

[0063] Among them, Q c,up,(t) represents the flow rate of the section in the river during period t, Q c,down,(t) represents the flow rate of the lower section of the river during period t, m 3 / s;L c,(t) is the water loss in the river or pipeline during period t, m 3 / s;

[0064] In addition to water level and flow constraints, river sections and water conservancy projects also have flow velocity constraints to prevent siltation and scouring. The constraint expressions are as follows:

[0065] v i,min ≤|v i，t |≤v i，max

[0066] Among them, v i，t is the flow velocity of the ith section at time t; v i,min and v i，max are the minimum and maximum flow velocities allowed at the i-th section at time t, respectively;

[0067] (c) Project operation constraints

[0068] 0≤q i ≤Q i，max

[0069] Among them, q i is the water diversion flow of the ith inlet; Q i，max is the maximum restricted flow rate of the i-th gate.

[0070] Furthermore, the NSGA-III algorithm is used to solve the river network dispatch optimization model, including:

[0071] (a) Initialize the population; Generate the initial population A according to the river network scheduling requirements t , that is, the first generation parent population, the number of individuals in the population is set to n, corresponding to n groups of scheduling schemes;

[0072] (b) Crossover and mutation: Perform crossover and mutation operations on the solutions in the parent population to generate a new child population. The crossover uses the simulated binary crossover algorithm, and the mutation uses the polynomial mutation algorithm. The calculation method is shown in the following formula:

[0073]

[0074] c′ kk =c k +δ(ul)

[0075] Among them, c' ii and c' jj are the new population individuals ii and jj generated by crossover; c' kk is the new population individual kk generated by mutation; c i 、c j and c k are individuals i, j and k of the parent population; β is the recombination coefficient corresponding to the simulated binary crossover algorithm; δ is the polynomial coefficient of variation; u is the upper limit of the decision variable, and l is the lower limit of the decision variable;

[0076] Thus, a new population of the same size of n is obtained; the parent generation and the new population are merged to obtain a candidate population of size 2n, from which n individuals need to be selected to generate a better population, i.e., the offspring population;

[0077] (c) Non-dominated sorting based on reference points; Dividing different non-dominated levels according to the Pareto dominance principle F Fi , divide the population into "non-dominated individuals > feasible individuals > infeasible individuals" for optimal iteration; select all individuals in higher levels in turn for retention, until a certain level cannot be fully accommodated, stop and abandon the subsequent levels; at this time, the selected individuals have It is still necessary to use reference points to filter out individuals; count(F Fq ) is the non-dominated layer F Fq Contains the number of individuals, Q is the number of levels fully placed;

[0078] (d) Reference point association and selection: According to the characteristics of the NSGA-III algorithm, a reference point mechanism is introduced; according to the correlation between each solution and the reference point, excellent solutions are selected to enter the next generation; first, the target space of the individual is normalized through the ideal point so that each target value is distributed in the interval [0,1]; then, according to the Euclidean distance between the normalized individual and the reference point, it is associated with the predefined reference point; finally, the reference point is selected in the order from sparse to dense, and the individuals associated with it are retained to ensure that the individuals are fully uniform in the target space until the required individual;

[0079] (e) Iterative update: Repeat steps (a) to (d) until the optimization is completed when the target value difference between the offspring and parent populations is less than 1% or the predetermined number of iterations is reached. Otherwise, the screened offspring population is used as the new parent population and all the above steps are repeated until the above convergence conditions are met, thereby obtaining a set of river network optimization scheduling schemes that do not dominate each other.

[0080] Based on the same inventive concept, a complex urban river network optimization and scheduling system of the present invention includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the complex urban river network optimization and scheduling method as described above.

[0081] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0082] The present invention integrates deep learning and the characteristics of urban river networks to construct a spatiotemporal graph matrix that reflects the dynamics of the river network and a spatiotemporal adjacency matrix that can simultaneously aggregate spatiotemporal information; based on the spatiotemporal graph neural network model, it can quickly predict the river network hydrological data within a certain time step in the future. According to the hydrological data predicted by the spatiotemporal graph neural network model, based on the determined objective function and constraints, a multi-objective river network scheduling optimization model is constructed, and the river network scheduling optimization model is solved by the NSGA-Ⅲ algorithm to output a scientific optimization scheduling plan.

[0083] In summary, the present invention has the advantages of high optimization scheduling quality and strong timeliness, and avoids the construction of complex river network hydrological and hydrodynamic physical models. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a flowchart of a complex urban river network optimization scheduling method provided by an embodiment of the present invention;

[0085] Figure 2 is a schematic diagram of a spatiotemporal adjacency matrix in an embodiment of the present invention;

[0086] Figure 3 It is the river network "four predictions" interactive page in the embodiment of the present invention;

[0087] Figure 4 This is the optimization decision scheduling solution page in the embodiment of the present invention. DETAILED DESCRIPTION

[0088] The present invention will be further described below in conjunction with the accompanying drawings.

[0089] Combination Figure 1 The embodiment of the present invention provides a method for predicting hydrological data of a complex urban river network, comprising:

[0090] (1) Collect and preprocess the hydrological data of urban river networks to form a hydrological dataset;

[0091] (101) Obtain hydrological data of urban river networks.

[0092] Obtain the hydrological data of existing hydrological stations and sluice pumps, deploy kettle software (open source data conversion software), extract the data of the corresponding stations through measuring point mapping, and synchronize the historical hydrological data into the database; for key river network nodes where water conservancy buildings and hydrological stations have not been built, observe the hydrological data through hydrological measuring equipment such as rain gauges, water level gauges, and flow meters, use 4G signal cards and remote communication acquisition unit equipment (RTU) for real-time data transmission, and the data center obtains the data and writes the data into the database.

[0093] (102) Preprocess the acquired urban river network hydrological data to form a hydrological data set.

[0094] Nodes were designed at key locations of the river network to collect statistics and fill in missing values. Since the magnitude of the data was sometimes quite different, the deviation standardization method was used to normalize the data. The conversion formula was:

[0095]

[0096] Among them, X * is the normalized data, which ranges from [0,1]; X is the original data; X max is the maximum value of the original data; X min is the minimum value of the original data.

[0097] (2) Establish a spatiotemporal graph neural network model;

[0098] The framework of the spatiotemporal graph neural network model is:

[0099] First, a spatiotemporal adjacency matrix is ​​generated according to the topological structure of the river network; then, the three-dimensional node spatiotemporal feature matrix is ​​input as input data into the fully connected layer, and the fully connected layer is used as the first layer to upgrade the node feature vector to enhance its expressive power; then, the spatiotemporal graph data after the feature vector upgrade is input into a set of spatiotemporal graph convolution combination modules for feature extraction, and the gating mechanism in the gated convolution is introduced, and the residual connection connected to the aggregation layer is introduced in each layer, so that the stacked graph convolution can achieve better results. After the spatiotemporal feature fusion extraction, the node features containing a large amount of spatiotemporal information are embedded in the global maximum pooling operation before entering the fully connected layer to form the final feature vector, which can eliminate redundant information and speed up the model processing speed. The final fully connected layer uses the softmax activation function to process the output and converts the output into a classification result, that is, the river network hydrological prediction data. Specifically:

[0100] (201) Taking hydrological stations and sluice pumps as nodes, and the water level, flow rate, flow velocity and other data of the nodes as the characteristic values ​​of the nodes, the data of a node n over a period of time can be expressed as a two-dimensional matrix N x , the feature matrices of all nodes are combined together to form a three-dimensional space-time graph matrix N.

[0101] (202) constructing a spatiotemporal adjacency matrix that can simultaneously aggregate spatiotemporal information on a node, combining the spatiotemporal adjacency matrix with the previously constructed three-dimensional spatiotemporal graph matrix, and connecting each node through the spatiotemporal adjacency matrix to form a complete spatiotemporal graph data sample;

[0102] Generate a spatial adjacency matrix A of size N×N based on the real topological structure of the river network S , which reflects the spatial connection status between river network nodes. Next, construct the spatiotemporal adjacency matrix A ST , change the information aggregation path, synchronously utilize the time and space characteristics of the measured values, and extract more effective spatiotemporal features. Spatiotemporal adjacency matrix A ST It consists of two sub-matrices, one of which is the above-mentioned spatial adjacency matrix A S , the other is the time adjacency matrix A T When considering processing s time steps in one graph convolution, we get a spatiotemporal adjacency matrix A of size sN×sN ST , A S Located on the diagonal, A T The nodes on both sides of the diagonal line are as follows: Figure 2 shown.

[0103] (203) The spatial domain graph convolution method is used to extract spatiotemporal fusion features from the spatiotemporal graph data samples. The attention mechanism is introduced to calculate the correlation between nodes, and the gating mechanism in the gated convolution is introduced. The residual connection connected to the aggregation layer is introduced in each layer, so that the stacked graph convolution can achieve better results; the spatiotemporal fusion features are subjected to global maximum pooling to form the final feature vector input into the fully connected layer. The fully connected layer uses the softmax activation function to output the classification result, namely the river network hydrological prediction data.

[0104] The spatial domain graph convolution method is used for graph convolution, through the feature matrix H l Matrix multiplication with the incidence matrix A aggregates information:

[0105] H l+1 =σ(AH l W l +b)

[0106] Among them, σ is the activation function; A is the adjacency matrix of the corresponding space-time graph; H l is the feature matrix of the lth layer; W lis the attention matrix of the lth layer; b is an additional bias used to fine-tune the result.

[0107] Considering that unconnected river network areas may also have strong correlations, the attention mechanism is introduced to calculate the correlation between nodes. Spatial attention can adaptively focus on the most relevant features based on the input data and can adaptively capture the spatial correlation between any nodes. The attention matrix W is involved in the calculation as shown in the following formula:

[0108] W=Oσ(Y l-1 Z1Z2(Z3Y l-1 ) T +b)

[0109] Among them, O is the mapping function, σ is the activation function, and Y l-1 is the output of the l-1th space-time block and the input of the lth space-time block; Z1, Z2, and Z3 are all learnable parameter matrices.

[0110] Connect multiple graph convolution modules to form a spatiotemporal graph convolution combined module to expand the spatial receptive field. In order to achieve better results for stacked graph convolution, the gating mechanism in gated convolution is introduced to reduce gradient diffusion and retain nonlinear capabilities. In the graph convolution process, the gated linear activation function is used as the activation function σ. The graph convolution process combined with the gating mechanism is described as:

[0111] H l+1 =(AH l W1+b1)⊙sigmoid(AH l W2+b2)

[0112] Where W1 and W2 are attention matrices; b1 and b2 are additional biases used to fine-tune the results; ⊙ represents the XOR operation; sigmoid(·) is a soft saturation activation function;

[0113] In addition to the gating mechanism, a residual connection connected to the aggregation layer is introduced in each layer, which is specifically described as:

[0114] H l+1 =F(H l ,A)+H l

[0115] Among them, H l represents the feature matrix of the lth layer; A represents the adjacency matrix of the spatiotemporal graph; F(H l ,A) indicates that H l The output feature matrix calculated by A, this operation usually includes steps such as convolution, nonlinear activation, and aggregation; H l+1 is the result after addition.

[0116] After the spatiotemporal feature fusion extraction, the node features containing a large amount of spatiotemporal information are embedded in the fully connected layer and then subjected to global maximum pooling to form the final feature vector. The pooling operation can eliminate redundant information and speed up the model processing. In this model, the maximum pooling is selected as the aggregation function to select the aggregation order that can contain the most information and reduce the required training parameters. After the maximum pooling, the s time steps processed by a graph convolution are cropped and merged into a two-dimensional node feature matrix containing the feature information of s time steps. The maximum pooling operation is expressed as:

[0117] h max-pooling =max(h 1 ,h 2 ,...,h L )

[0118] Among them, h L Represents the eigenvalue of the Lth input.

[0119] The spatiotemporal fusion features of the entire spatiotemporal graph data sample can be extracted by combining s spatiotemporal graph convolutional combination modules into a group in a parallel structure and splicing the output features in sequence.

[0120] Finally, the fully connected layer uses the softmax activation function to process the output and convert the output into hydrological prediction data for each node.

[0121] (3) The hydrological data set is used to train and optimize the spatiotemporal graph neural network model, and the trained spatiotemporal graph neural network model is used to predict hydrological data.

[0122] The hydrological data set is divided into a training set and a test set. When the training set is used to input the model for training, the weight of each layer of the model is initialized first, and then the data is input. The output of the model is calculated forward, and then the loss is calculated according to the loss function. The loss function uses the mean square error MSE. The specific formula is:

[0123]

[0124] Where n is the number of samples, y i is the actual label, is the model prediction.

[0125] After calculating the loss, the gradient is back-propagated to update the weights of each layer. The model is fit and optimized through multiple iterations of gradient descent. The Adam algorithm is used to update the weights according to these gradients. The regularization method is used to prevent overfitting, and finally a trained spatiotemporal graph neural network model is obtained.

[0126] Combination Figure 1The embodiment of the present invention also provides a method for optimizing the scheduling of a complex urban river network, including: constructing a river network scheduling optimization model based on the determined objective function and constraints according to the hydrological prediction data obtained by the method for predicting hydrological data of a complex urban river network described in the embodiment of the present invention; and solving the river network scheduling optimization model using the NSGA-III algorithm to obtain a river network scheduling plan.

[0127] 1) Construct the objective function, which includes ecological flow rate and ecological water level.

[0128] (a) Ecological flow rate

[0129] v≥max(v1,v2)

[0130] Among them, v is the ecological flow rate; v1 is the flow rate target to ensure the flow of the river channel. The RVA calculation is performed using the variation range method. The RVA threshold usually uses the values ​​corresponding to the 75% and 25% probability of occurrence of the hydrological evaluation index in the long sequence year as the upper and lower limits. This threshold range is considered to be the ecological flow variation range for the river ecosystem to maintain health. The determination of the ecological flow rate can refer to this calculation method. The calculation formula is:

[0131]

[0132] in, is the multi-year average monthly mean of flow velocity; v 上限 v is the value corresponding to 75% probability of flow rate occurrence; 下限 It is the value corresponding to 25% probability of flow rate occurrence.

[0133] v2 is the velocity required for the survival environment of fish in urban rivers and the velocity when algae growth is inhibited. The velocity to maintain the function of aquatic habitat is 0.1 m / s.

[0134] When the ecological flow rate guarantee rate of a river reaches 50%, it is considered to meet the ecological flow rate requirements. The ecological flow rate guarantee rate of a river refers to the proportion of the duration greater than the ecological flow rate to the total duration;

[0135] (b) Ecological water level

[0136] H4>H≥max(H1,H2,H3)

[0137] Among them, H1 is the ecological water level required to maintain river connectivity, which is also calculated using the RVA method; H2 is the ecological water depth to maintain fish habitats, which needs to consider the survival depth required by fish at different growth stages, and the water depth required for fish to turn vertically. Under normal circumstances, fish are suitable for growing in a water depth of 2 to 3 times their body length; H3 is the ecological water level required to maintain good landscape and entertainment factors, and the calculation formula is:

[0138] H3=max(H通航 ,H 灌溉 ,H 亲水 )

[0139] Among them, H 通航 The lowest river water level required for navigation. Normally, the water depth required for small cruise ships is 1.5m. 亲水 The water level of the river is suitable for the hydrophilic leisure platform. Studies have shown that when the hydrophilic platform is within 0.5m from the river surface, the hydrophilic effect of the platform is better; H 灌溉 Minimum water level required for irrigation on both sides of the river:

[0140] H 灌溉 =min(H 灌溉1 ,H 灌溉2 ,…,H 灌溉j )

[0141] Among them, H 灌溉j is the designed minimum water intake level of the jth water intake structure;

[0142] H4 is the river water level that should not be exceeded before the flood season for flood control and municipal drainage. The calculation formula is:

[0143] H4=min(H 防洪 ,H 市政 )

[0144] Among them, H 防洪 The flood control level stipulated by the water management department; 市政 To ensure stable water level of municipal drainage system:

[0145] H 市政 =min(H 排1 ,H 排2 ,…,H 排m )

[0146] Among them, H 排m is the water level elevation of the mth drainage outlet;

[0147] When the ecological water level protection rate of the river reaches 50%, it is deemed to meet the ecological water level requirements. At the same time, the water level limit for flood control and drainage shall not be exceeded at any time.

[0148] 2) Determine the constraints, which include water balance constraints of sluice pumps, flow and velocity constraints of various levels of rivers, and project operation constraints.

[0149] (a) Water balance constraints of sluice pumps

[0150] Q b,in,(k,t) =Q b,out,(k,t)

[0151] Among them, Qb,in,(k,t) represents the water diversion flow of the sluice pump during period t, Q b,out,(k,t) Indicates the water flow rate of the sluice pump during period t, m 3 / s;

[0152] (b) Constraints on flow and velocity at all levels of rivers

[0153] Q c,up,(t) =Q c,down,(t) +L c,(t)

[0154] Among them, Q c,up,(t) represents the flow rate of the section in the river during period t, Q c,down,(t) represents the flow rate of the lower section of the river during period t, m 3 / s;L c,(t) is the water loss in the river or pipeline during period t, m 3 / s;

[0155] In addition to water level and flow constraints, river sections and water conservancy projects also have flow velocity constraints to prevent siltation and scouring. The constraint expressions are as follows:

[0156] v i,min ≤|v i，t |≤v i，max

[0157] Among them, v i，t is the flow velocity at the i-th section at time t. In actual calculation, the flow velocity has a direction (positive or negative), so here we only compare the magnitude; v i,min and v i，max are the minimum and maximum flow velocities allowed at the i-th section at time t, respectively;

[0158] (c) Project operation constraints

[0159] 0≤q i ≤Q i，max

[0160] Among them, q i is the water diversion flow of the ith inlet; Q i，max is the maximum restricted flow rate of the i-th gate.

[0161] 3) Using NSGA-Ⅲ algorithm to solve the river network dispatch optimization model

[0162] (a) Initialize the population; Generate the initial population A according to the river network scheduling requirements t , that is, the first generation parent population, the number of individuals in the population is set to n, corresponding to n groups of scheduling schemes;

[0163] (b) Crossover and mutation: Perform crossover and mutation operations on the solutions in the parent population to generate a new child population. The crossover uses the simulated binary crossover algorithm, and the mutation uses the polynomial mutation algorithm. The calculation method is shown in the following formula:

[0164]

[0165] c′ kk =c k +δ(ul)

[0166] Among them, c' ii and c' jj are the new population individuals ii and jj generated by crossover; c' kk is the new population individual kk generated by mutation; c i 、c j and c k are individuals i, j and k of the parent population; β is the recombination coefficient corresponding to the simulated binary crossover algorithm; δ is the polynomial coefficient of variation; u is the upper limit of the decision variable, and l is the lower limit of the decision variable;

[0167] Thus, a new population of the same size of n is obtained; the parent generation and the new population are merged to obtain a candidate population of size 2n, from which n individuals need to be selected to generate a better population, i.e., the offspring population;

[0168] (c) Non-dominated sorting based on reference points and: Dividing different non-dominated levels according to the Pareto dominance principle F Fi , divide the population into "non-dominated individuals > feasible individuals > infeasible individuals" for optimal iteration; select all individuals in higher levels in turn for retention, until a certain level cannot be fully accommodated, stop and abandon the subsequent levels; at this time, the selected individuals have It is still necessary to use reference points to filter out (Here “-” indicates a minus sign) individuals; count(F Fq ) is the non-dominated layer F Fq Contains the number of individuals, Q is the number of levels fully placed;

[0169] (d) Reference point association and selection: According to the characteristics of the NSGA-III algorithm, a reference point mechanism is introduced; according to the association degree between each solution and the reference point (such as distance, congestion, etc.), excellent solutions are selected to enter the next generation; first, the target space of the individual is normalized through the ideal point so that each target value is distributed in the interval [0,1]; then, according to the Euclidean distance between the normalized individual and the reference point, it is associated with the predefined reference point; finally, the reference point is selected in the order from sparse to dense, and the individuals associated with it are retained to ensure that the individuals are fully uniform in the target space until the required individual;

[0170] (e) Iterative update: Repeat steps (a) to (d) until the optimization is completed when the target value difference between the offspring and parent populations is less than 1% or the predetermined number of iterations is reached. Otherwise, the screened offspring population is used as the new parent population and all the above steps are repeated until the above convergence conditions are met, thereby obtaining a set of river network optimization scheduling schemes that do not dominate each other.

[0171] An embodiment of the present invention also provides a complex urban river network optimization and scheduling system, including a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the electronic device implements the steps of the method described in the above embodiment and can achieve the same technical effect as the above method.

[0172] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, the memory, such program modules including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include the implementation of a network environment. The program modules typically perform the functions and / or methods in the embodiments described herein.

[0173] The processor executes various functional applications and data processing by running the programs stored in the memory, such as implementing the method provided in the embodiment of the present invention.

[0174] Based on the spatiotemporal graph neural network model and river network scheduling optimization model described in the embodiments of the present invention, the following applications can be realized. Specifically:

[0175] Based on the SCADA system, the four pre-calculations and scheduling decision-making system of "forecast, warning, rehearsal, and emergency plan": Based on the spatiotemporal graph neural network model and the water conservancy project SCADA system, on-site monitoring and control parameter measurement points are constructed, and data samples are extracted and trained based on the measurement points. The trained spatiotemporal graph neural network model is obtained through parameter adjustment, and the input and output of the measurement points are visualized, and the output parameter set GIS (geographic information system) is used to display the "four pre-calculations" of scheduling. At the same time, according to the "four pre-calculations" results of the urban river network, the optimization solution is obtained through the river network scheduling optimization model to obtain the optimal scheduling plan corresponding to the river network. The specific page is as follows: Figure 3 and Figure 4As shown, the urban river network water diversion scheduling, water situation forecasting and prediction, and optimal water scheduling and allocation calculation are realized, providing a reliable scheduling decision-making basis for water conservancy workers to issue scheduling instructions.

Claims

1. A method for predicting hydrological data of a complex urban river network, characterized in that: include: (1) Collect and preprocess the hydrological data of urban river networks to form a hydrological dataset; (2) Establish a spatiotemporal graph neural network model; (201) Taking hydrological stations and sluice pumps as nodes, and the water level, flow rate, and velocity data of the nodes as the characteristic values ​​of the nodes, the characteristic matrices of all nodes within a period of time are combined together to form a three-dimensional space-time graph matrix N; (202) constructing a spatiotemporal adjacency matrix that can simultaneously aggregate spatiotemporal information on a node, combining the spatiotemporal adjacency matrix with the three-dimensional spatiotemporal graph matrix, and connecting each node through the spatiotemporal adjacency matrix to form a complete spatiotemporal graph data sample; (203) The spatial domain graph convolution method is used to extract spatiotemporal fusion features from the spatiotemporal graph data samples, an attention mechanism is introduced to calculate the correlation between nodes, and a gating mechanism in the gated convolution is introduced, and a residual connection connected to the aggregation layer is introduced in each layer; the spatiotemporal fusion features are subjected to global maximum pooling to form the final feature vector and input into the fully connected layer, which uses the softmax activation function to output the classification results, namely the river network hydrological prediction data; (3) The hydrological data set is used to train and optimize the spatiotemporal graph neural network model, and the trained spatiotemporal graph neural network model is used to predict hydrological data.

2. The method for predicting hydrological data of complex urban river networks according to claim 1, characterized in that: Step (1) comprises: (101) Obtaining urban river network hydrological data Obtain hydrological data from existing hydrological stations and sluice pumps; for key river network nodes where no water conservancy buildings or hydrological stations have been built, observe hydrological data through hydrological measurement equipment; (102) Preprocess the acquired urban river network hydrological data to form a hydrological data set Nodes were designed at key locations of the river network, the collected data were counted, and missing values ​​were filled in. The data were normalized using the deviation standardization method, and the conversion formula was: Among them, X * is the normalized data, which ranges from [0,1]; X is the original data; X max is the maximum value of the original data; X min is the minimum value of the original data.

3. The method for predicting hydrological data of complex urban river networks according to claim 1 is characterized in that: Step (202) includes: generating a spatial adjacency matrix A of size N×N according to the real topological structure of the river network S , then construct the spatiotemporal adjacency matrix A ST , the spatiotemporal adjacency matrix A ST It consists of two sub-matrices, including the spatial adjacency matrix A S and the time adjacency matrix A T ; When considering processing s time steps in a graph convolution, the spatiotemporal adjacency matrix A of size sN×sN is obtained ST , A S Located on the diagonal, A T The nodes are located on both sides of the diagonal line.

4. The method for predicting hydrological data of complex urban river networks according to claim 1, characterized in that: Step (203) includes: using the spatial domain graph convolution method to perform graph convolution, through the feature matrix H l Matrix multiplication with the incidence matrix A aggregates information: H l+1 =σ(AH l W l +b) Among them, σ is the activation function; A is the adjacency matrix of the corresponding space-time graph; H l is the feature matrix of the lth layer; W l is the attention matrix of the lth layer; b is an additional bias used to fine-tune the results; The attention mechanism is introduced to calculate the correlation between nodes. The attention matrix W is involved in the calculation as shown in the following formula: W=Oσ(Y l-1 Z1Z2(Z3Y l-1 ) T +b) Among them, O is the mapping function, Y l-1 is the output of the l-1th space-time block and the input of the lth space-time block; Z1, Z2, and Z3 are all learnable parameter matrices; Multiple graph convolution modules are connected to form a spatiotemporal graph convolution combination module to expand the spatial receptive field. In the graph convolution process, a gated linear activation function is used as the activation function σ. The graph convolution process combined with the gating mechanism is described as: <h2 style=";text-align:left;direction:ltr">H<h2 style=";text-align:left;direction:ltr"> l+1 <h2 style=";text-align:left;direction:ltr"> (AH)<h2 style=";text-align:left;direction:ltr"> l <h2 style=";text-align:left;direction:ltr"> W1+b1)⊙sigmoid(AH<h2 style=";text-align:left;direction:ltr"> l <h2 style=";text-align:left;direction:ltr"> W2+b2) Where W1 and W2 are attention matrices; b1 and b2 are additional biases used to fine-tune the results; ⊙ represents the XOR operation; sigmoid(·) is a soft saturation activation function; In addition to the gating mechanism, a residual connection connected to the aggregation layer is introduced in each layer, which is specifically described as: H l+1 =F(H l ,A)+H l Among them, F(H l ,A) indicates that H l The output feature matrix calculated by and A; H l+1 is the result after addition.

5. The method for predicting hydrological data of complex urban river networks according to claim 1, characterized in that: Step (3) includes: using mean square error as the loss function, calculating the gradient back-propagation to update the weights of each layer after calculating the loss, fitting the model and reaching the optimal value through multiple iterations of gradient descent, using the Adam algorithm to update the weights according to these gradients, and using regularization methods to prevent overfitting.

6. A complex urban river network optimization scheduling method, characterized in that: include: According to the hydrological prediction data obtained by the complex urban river network hydrological data prediction method described in any one of claims 1 to 5, a river network scheduling optimization model is constructed based on the determined objective function and constraints; the NSGA-III algorithm is used to solve the river network scheduling optimization model to obtain a river network scheduling plan.

7. The complex urban river network optimization scheduling method according to claim 6 is characterized in that: The objective function includes ecological flow rate and ecological water level: (a) Ecological flow rate v≥max(v1,v2) Among them, v is the ecological flow rate; v1 is the velocity target to ensure the flow of the river, and is calculated using the RVA method using the following formula: in, is the multi-year average monthly mean of flow velocity, v 上限 and v 下限 The values ​​corresponding to 75% and 25% of the probability of occurrence of hydrological evaluation indicators in long-sequence years are selected respectively; v2 is the flow velocity required by the living environment of urban river fish and the flow velocity when algae growth is inhibited; When the ecological flow rate guarantee rate of a river reaches 50%, it is considered to meet the ecological flow rate requirements. The ecological flow rate guarantee rate of a river refers to the proportion of the duration greater than the ecological flow rate to the total duration; (b) Ecological water level H4>H≥max(H1,H2,H3) Among them, H1 is the ecological water level required to maintain river connectivity, which is also calculated using the RVA method; H2 is the ecological water depth to maintain fish habitats, which needs to consider the survival depth required by fish at different growth stages, and the water depth required for fish to turn vertically; H3 is the ecological water level required to maintain good landscape entertainment factors, and the calculation formula is: H3=max(H 通航 ,H 灌溉 ,H 亲水 ) Among them, H 通航 The lowest river water level required for navigation; H 亲水 The water level of the river is suitable for the water-friendly leisure activity platform; H 灌溉 Minimum water level required for irrigation on both sides of the river: H 灌溉 =min(H 灌溉1 ,H 灌溉2 ,…,H 灌溉j ) Among them, H 灌溉j is the designed minimum water intake level of the jth water intake structure; H4 is the river water level that should not be exceeded before the flood season for flood control and municipal drainage. The calculation formula is: H4=min(H 防洪 ,H 市政 ) Among them, H 防洪 The flood control level stipulated by the water management department; 市政 To ensure stable water level of municipal drainage system: H 市政 =min(H 排1 ,H 排2 ,…,H 排m ) Among them, H 排m is the water level elevation of the mth drainage outlet; When the ecological water level protection rate of the river reaches 50%, it is deemed to meet the ecological water level requirements. At the same time, the water level limit for flood control and drainage shall not be exceeded at any time.

8. The complex urban river network optimization scheduling method according to claim 7 is characterized in that: The constraints include water balance constraints of sluice pumps, flow and velocity constraints of various levels of rivers, and engineering operation constraints: (a) Water balance constraints of sluice pumps Q b,in,(k,t) =Q b,out,(k,t) Among them, Q b,in,(k,t) represents the water diversion flow of the sluice pump during period t, Q b,out,(k,t) Indicates the water flow rate of the sluice pump during period t, m 3 / s; (b) Constraints on flow and velocity at all levels of rivers Q c,up,(t) =Q c,down,(t) +L c,(t) Among them, Q c,up,(t) represents the flow rate of the section in the river during period t, Q c,down,(t) represents the flow rate of the lower section of the river during period t, m 3 / s;L c,(t) is the water loss in the river or pipeline during period t, m 3 / s; In addition to water level and flow constraints, river sections and water conservancy projects also have flow velocity constraints to prevent siltation and scouring. The constraint expressions are as follows: v i,min ≤|v i，t |≤v i，max Among them, v i，t is the flow velocity of the ith section at time t; v i,min and v i，max are the minimum and maximum flow velocities allowed at the i-th section at time t, respectively; (c) Project operation constraints 0≤q i ≤Q i，max Among them, q i is the water diversion flow of the ith inlet; Q i，max is the maximum restricted flow rate of the i-th gate.

9. The complex urban river network optimization scheduling method according to claim 8 is characterized in that: The NSGA-Ⅲ algorithm is used to solve the river network dispatch optimization model, including: (a) Initialize the population; Generate the initial population A according to the river network scheduling requirements t , that is, the first generation parent population, the number of individuals in the population is set to n, corresponding to n groups of scheduling schemes; (b) Crossover and mutation: Perform crossover and mutation operations on the solutions in the parent population to generate a new child population. The crossover uses the simulated binary crossover algorithm, and the mutation uses the polynomial mutation algorithm. The calculation method is shown in the following formula: c′ kk =c k +δ(u-l) Among them, c' ii and c' jj are the new population individuals ii and jj generated by crossover; c' kk is the new population individual kk generated by mutation; c i 、c j and c k are individuals i, j and k of the parent population; β is the recombination coefficient corresponding to the simulated binary crossover algorithm; δ is the polynomial coefficient of variation; u is the upper limit of the decision variable, and l is the lower limit of the decision variable; Thus, a new population of the same size of n is obtained; the parent generation and the new population are merged to obtain a candidate population of size 2n, from which n individuals need to be selected to generate a better population, i.e., the offspring population; (c) Non-dominated sorting based on reference points; Dividing different non-dominated levels according to the Pareto dominance principle F Fi , divide the population into "non-dominated individuals > feasible individuals > infeasible individuals" for optimal iteration; select all individuals in higher levels in turn for retention, until a certain level cannot be fully accommodated, stop and abandon the subsequent levels; at this time, the selected individuals have It is still necessary to use reference points to filter out individuals; count(F F q) is the non-dominated layer F Fq Contains the number of individuals, Q is the number of levels fully placed; (d) Reference point association and selection: According to the characteristics of the NSGA-III algorithm, a reference point mechanism is introduced; according to the correlation between each solution and the reference point, excellent solutions are selected to enter the next generation; first, the target space of the individual is normalized through the ideal point so that each target value is distributed in the interval [0,1]; then, according to the Euclidean distance between the normalized individual and the reference point, it is associated with the predefined reference point; finally, the reference point is selected in the order from sparse to dense, and the individuals associated with it are retained to ensure that the individuals are fully uniform in the target space until the required individual; (e) Iterative update: Repeat steps (a) to (d) until the optimization is completed when the target value difference between the offspring and parent populations is less than 1% or the predetermined number of iterations is reached. Otherwise, the screened offspring population is used as the new parent population and all the above steps are repeated until the above convergence conditions are met, thereby obtaining a set of river network optimization scheduling schemes that do not dominate each other.

10. A complex urban river network optimization dispatching system, characterized in that: The device comprises a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the device implements the steps of the complex urban river network optimization scheduling method as described in any one of claims 6 to 9.

Images