Runoff prediction method and device and storage medium

Abstract

The invention discloses a runoff prediction method, a runoff prediction device and a storage medium. The runoff prediction method comprises the following steps: calculating dynamic response characteristics of each grid unit at each moment by using meteorological input; extracting static influence characteristics of the static attributes on the runoff error; extracting features reflecting time and space influences at the same time, and obtaining an error correction value generated based on meteorological input and underlying surface information; constructing a reservoir scheduling substitution network, and outputting a correction component of the reservoir scheduling to the runoff; and constructing a loss function, superposing the correction component and the output of the preliminary hydrological model, and updating parameters of each module through end-to-end back propagation to obtain runoff prediction. According to the runoff prediction method, the influence of dynamic changes such as rainfall and air temperature, underlying surface distribution and reservoir regulation on drainage basin runoff is fully captured, and the precision and stability of runoff simulation are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological forecasting, and particularly relates to a method for predicting runoff by coupling hydrological system response and reservoir operation alternative network. Background Art

[0002] Hydrological forecasting and reservoir operation play crucial roles in water resources management and flood control and disaster reduction. Under the influence of climate change and human activities, there are uncertainties in climate elements and underlying surface characteristics, and there are often errors in runoff simulation results. In addition, traditional hydrological models are difficult to quantify the impact of complex reservoir operation on the runoff formation process, reducing the accuracy of runoff forecasting. Therefore, comprehensively considering the superposition effect of hydrological system response and reservoir operation on runoff error is of great significance for improving runoff prediction accuracy and basin water management. Along with the development of machine learning models, data-driven methods can learn the complex non-linear relationships of hydrological systems through large-scale historical data, correct runoff errors and improve prediction accuracy. However, existing machine learning methods often only model the forecasting model or reservoir operation separately, and fail to comprehensively consider the error sources of runoff forecasting and their propagation mechanisms, resulting in limited runoff prediction levels. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a runoff prediction method, device and storage medium that effectively couple hydrological system response and reservoir operation alternative network.

[0004] To achieve the above object, the present invention specifically adopts the following technical solutions: The present invention first provides a runoff prediction method, including the following steps: Step S1: Use meteorological inputs to construct a dynamic system response curve, and calculate the dynamic response characteristics of each time and each grid cell, where the meteorological inputs include precipitation and temperature; Step S2: Extract the static influence characteristics of runoff error from the basin underlying surface data through convolution and non-linear transformation; Step S3: Use a neural network to align the dynamic response characteristics and static influence characteristics inputs, extract the characteristics that simultaneously reflect temporal and spatial influences, and obtain the error correction amount generated based on meteorological inputs and underlying surface information; Step S4: Combine the reservoir operation status and scheduling requirements to construct a reservoir operation alternative network, and output the correction component of reservoir operation on runoff; Step S5: Collect runoff forecast and measured data to construct a loss function, superimpose the above correction component and the output of the preliminary hydrological model, and simultaneously update the neural network model parameters through end-to-end backpropagation to obtain the runoff prediction amount after error correction.

[0005] Runoff prediction amount: The sources of runoff error are attributed to the uncertainties of dynamic meteorological input data and static basin underlying surface data, as well as the impact of reservoir operation on runoff not being considered. The runoff prediction amount is equal to the sum of the simulated values of the hydrological model after restoring these three parts of errors. The uncertainty of meteorological data transmits errors through the dynamic system response curve. The uncertainty of the underlying surface data can provide static feature inputs for the neural network model. Reservoir operation serves as an alternative model output for the runoff correction component of the physical neural network.

[0006] In step S5, the specific algorithm for calculating the runoff prediction amount within the calculation period is as follows: In the formula, is the runoff prediction amount (m³ / s) after error correction within the period; is for the period the simulated flow rate (m³ / s) of the original hydrological model; the period the error correction amount generated based on meteorological input and underlying surface information; is for the period the correction amount of the impact of reservoir operation on runoff. For the errors caused by meteorological input and underlying surface information, runoff generation is significantly affected by the lag effect of precipitation data, that is, the precipitation errors in the previous few periods will be transmitted to the runoff error in the current period through the system response curve; at the same time, the underlying surface data provides feature inputs for each grid as static attributes.

[0007] Therefore, the calculation of is shown in formula (2). In the formula, is the dynamic and static element fusion module; is the dynamic response feature;

[0008] The specific method for convolving and fusing the dynamic response feature and the static feature is as follows: In the formula, is the output feature map; is the index of the output channel; is the activation function; is for the period at the grid the concatenated feature vector; is the input channel index, that is, the total number after concatenating the dynamic and static feature inputs; and are the number of channels of the dynamic feature input data and the static feature input data respectively; is the learnable weight of the 1×1 convolution kernel in the channel dimension; is the bias of the convolution kernel.

[0009] In step S3, the calculation period In the grid The specific method for splicing the feature vectors is as follows: In the formula, denotes splicing the feature data in the channel dimension.

[0010] In step S1, the dynamic response feature The specific calculation method is as follows: In the formula, is the considered lag order, that is, in the period convolve the data of the past periods; and are the period precipitation and temperature at the grid scale, respectively; and are the weight coefficients of precipitation and temperature data at the grid scale, respectively.

[0011] In step S2, the specific method for calculating the basin static characteristics is as follows: In the formula, is the static attribute of the grid , including soil type, vegetation type, and catchment area; represents a trainable neural network used to extract the influence of static attributes on runoff error.

[0012] In step S3, the specific method for calculating the error correction amount generated based on meteorological input and underlying surface information is as follows: In the formula, is the spatio-temporal correction weight, which can reflect the propagation mechanism of spatial heterogeneity, time lag effect, and catchment process on error in the hydrological process within the basin to a certain extent as a parameter; is the output feature map, is the index of the output channel. The spatio-temporal correction weight is calculated as follows: In the formula, is the catchment area corresponding to this grid point.

[0013] In step S4, the specific algorithm for constructing the reservoir operation alternative network is as follows: In the formula, the input layer includes the reservoir water level within the period, the reservoir inflow and the reservoir storage , and the reservoir storage is obtained by the water level-storage relationship curve; the output layer is , through the water balance equation and the constraints of the outflow and water level limits.

[0014] In step S5, the specific method for calculating the loss function is as follows: In the formula, is the loss function (root mean square error); is the total number of time steps for calculating the error; is the observed runoff.

[0015] The present invention also provides a runoff prediction device, including a processor and a memory; a program or instruction is stored in the memory, and the program or instruction is loaded and executed by the processor to implement the steps of the runoff prediction method.

[0016] The present invention also provides a computer-readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the runoff prediction method are implemented.

[0017] Advantages of the present invention: The present invention proposes a runoff prediction method that couples the response of the hydrological system and reservoir operation. First, a dynamic system response curve is constructed using meteorological inputs such as precipitation and temperature to calculate the dynamic response characteristics at each time and each grid cell. Then, the influence characteristics of the static attributes of the basin underlying surface on runoff error are extracted through convolution and non-linear transformation. Next, a neural network is used to align the dynamic and static feature inputs to extract features that simultaneously reflect temporal and spatial influences and generate spatio-temporal correction weights. After that, a reservoir operation substitution network is constructed in combination with the reservoir operation status and scheduling requirements to output the correction component of reservoir operation on runoff. Finally, a loss function is constructed by collecting runoff forecasts and measured data, and the above correction component is superimposed on the output of the preliminary hydrological model. The parameters of each module are updated simultaneously through end-to-end backpropagation to obtain the runoff prediction. The present invention proposes a runoff prediction method that couples the response of the hydrological system and the reservoir operation substitution network, considering the superposition effects of dynamic system response, basin static attributes, and reservoir operation on runoff error, solving the problem that current runoff error correction methods are difficult to quantify various error sources and their propagation mechanisms, and effectively improving the accuracy of hydrological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic flow chart of the runoff prediction method that couples the response of the hydrological system and reservoir operation provided by the present invention; Figure 2 is a distribution map of basin dynamic response characteristics in a specific embodiment; Figure 3 is a distribution map of basin static characteristics in a specific embodiment; Figure 4 is the process of the reservoir operation water level process line in the basin in a specific embodiment; Figure 5It is the daily runoff correction result of the runoff prediction method that couples the hydrological system response and reservoir operation in the specific embodiment. Detailed implementation mode

[0019] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0020] It should be understood that the specific implementation mode described herein is only used to explain the present invention and is not used to limit the present invention.

[0021] Embodiment 1 As Figure 1 shown, the present invention provides a runoff prediction method that couples the hydrological system response and reservoir operation, including the following steps: Step S1, constructing a dynamic system response curve by using meteorological inputs such as precipitation and temperature, and calculating the dynamic response characteristics of each moment and each grid cell. The specific steps include: Step S101, constructing a dynamic system response curve by using meteorological inputs such as precipitation and temperature; Step S102, calculating the dynamic response characteristics of each moment and each grid cell: In the formula, is the dynamic response characteristic of grid in time period ; and are the weight coefficients of precipitation and temperature data at the grid scale when the lag days are respectively; is the considered lag order, that is, convolution calculation is performed on the data of the past time periods in time period ; and are the precipitation and temperature of grid when the lag days are respectively.

[0022] Obtain the precipitation and temperature grid data of the Jin'anqiao Basin in the middle reaches of the Jinsha River Basin from 2013 to 2020. Take each grid cell of the meteorological data as a calculation unit, construct the system response relationship between the runoff error component and the meteorological input error, and calculate the dynamic response characteristics of each grid cell by using convolution transformation, as Figure 2 shown.

[0023] Step S2, extracting the static influence characteristics of the static attributes of the basin underlying surface data on the runoff error through convolution and nonlinear transformation. The specific steps include: Step S201, extracting the static attributes of the basin such as soil type, vegetation type, and catchment area; Step S202: Map the basin static attributes to the impact feature space of runoff error through convolution and non-linear transformation to obtain the basin static features: In the formula, is the basin static feature of grid ; is the static attribute of grid ; represents a trainable neural network used to extract the basin static features of the impact of static attributes on runoff error. is a known network, and the output of the network is the encoded soil-vegetation-confluence area composite attribute vector, that is, the static feature .

[0024] Download the soil type and vegetation type data, use the GIS spatial analysis tool to obtain the static attributes of the Jin'anqiao Basin, and map the static attributes of the Jin'anqiao Basin to the impact feature space of runoff error through convolution and non-linear transformation, as Figure 3 shown.

[0025] Step S3: Use the neural network to align the dynamic features and static feature inputs, extract the features that simultaneously reflect the temporal and spatial impacts, and obtain the error correction amount generated based on the meteorological input and underlying surface information.

[0026] Step S301: Use the neural network to align the dynamic features and static feature inputs: In the formula, is the concatenated feature vector after alignment and concatenation, , indicates that the concatenated feature map is a three-dimensional tensor with a spatial size of and a number of channels of ; is the input channel index, that is, the total number after the dynamic and static feature inputs are concatenated; indicates concatenating the feature data in the channel dimension; and are the number of channels of the dynamic feature input data and the static feature input data respectively; is the longitudinal number of the grid, is the transverse number of the grid.

[0027] Step S302: Use convolution to fuse the dynamic and static features to obtain the features that can simultaneously reflect the temporal and spatial impacts: In the formula, is the output feature map; is the index of the output channel; is the activation function; is the learnable weight of the 1×1 convolution kernel in the channel dimension; Is the bias of the convolution kernel.

[0028] Step S303: Define the spatio-temporal correction weight and calculate the error correction amount generated based on the meteorological input and the underlying surface information; Among them, the spatio-temporal correction weight is defined as: In the formula, is the spatio-temporal correction weight, which reflects the spatial heterogeneity involved in the hydrological process within the basin through the catchment area of each grid cell and the propagation and contribution of the confluence process to the error; The catchment area corresponding to this grid point; Based on the defined spatio-temporal correction weight , deduce the error correction component generated by the meteorological input and the underlying surface information: In the formula, is the error correction component generated by the meteorological input and the underlying surface information.

[0029] Determine the spatio-temporal correction weight of each grid cell according to the proportion of the confluence area of each grid cell in the Jin'anqiao Basin, and deduce the error correction component generated by the meteorological input and the underlying surface information.

[0030] Step S4: Combine the reservoir operation status and the scheduling requirements to construct a reservoir scheduling alternative network and output the correction component of the reservoir scheduling to the runoff.

[0031] Step S401: Construct a water balance equation to quantify the reservoir operation status and the scheduling requirements: In the formula, is the inflow during the time period ; is the outflow during the time period ; is the reservoir storage during the time period , obtained through the water level-storage relationship; is the reservoir water level during the time period , is the functional relationship between the water level and the storage, used to calculate the reservoir storage corresponding to this water level; Equation (9) and Equation (10) are the outflow constraint and the water level limit constraint respectively.

[0032] Step S402: Combine the reservoir operation status and the scheduling requirements to construct a reservoir scheduling alternative network and output the correction component of the reservoir scheduling to the runoff: In the formula, is the reservoir scheduling alternative model constructed based on the neural network, is the trainable parameter; is the output layer, that is, the correction component of the reservoir scheduling to the runoff, and the outflow range is restricted by the activation function: In the formula, They are the weights of the first layer, the second layer, and the output layer of the reservoir surrogate network respectively; The biases of the corresponding reservoir surrogate network; is the activation function of the reservoir surrogate network; and are intermediate hidden layer variables, representing the output results of the neurons in the first layer and the second layer respectively.

[0033] Combined with the operation status and scheduling requirements of Liyuan Reservoir, the downstream water level process line of Liyuan Reservoir in the upper reaches of Jin'anqiao Basin is obtained, as Figure 4 shown.

[0034] Step S5, collect runoff forecast and measured data to construct a loss function, superimpose the above correction components and the output of the preliminary hydrological model, and update the parameters of each module simultaneously through end-to-end backpropagation to obtain the runoff prediction.

[0035] Step S501, collect runoff forecast and measured data to construct a loss function: In the formula, is the loss function (mean square error); is the total number of time steps for calculating the error; Observed runoff (m³ / s); is the runoff prediction after error correction within the time period.

[0036] Step S502, superimpose the system response correction component and the reservoir operation correction component with the output of the preliminary hydrological model, and update the network model parameters simultaneously through end-to-end backpropagation to obtain the runoff prediction: In the formula, is the runoff prediction after error correction within the time period; is the time period The simulated flow of the original hydrological model under; Time period The error correction amount generated based on meteorological input and underlying surface information under; is the time period The correction amount of the impact of reservoir operation on runoff under.

[0037] Taking 2014 - 2019 as the calibration period, 2020 - 2022 as the verification period, and 2023 - 2024 as the test period, the effect of the corrected runoff in Jin'anqiao Basin is as Figure 5As shown in the figure. According to Step 1, the dynamic response characteristics of each moment and each grid unit in the Jin'anqiao Basin are calculated using precipitation and temperature. Then, the influence characteristics of the underlying surface information of the extracted basin on the runoff error are obtained. Next, the dynamic and static data are fused to extract the characteristics that reflect both temporal and spatial influences, and the spatio-temporal correction weights are generated. Then, considering the operation status and scheduling requirements of the Liyuan Hydropower Station Reservoir in the upper reaches of the Jin'anqiao Basin, a reservoir scheduling alternative network is constructed, and the correction component of the reservoir scheduling on the runoff is output. Finally, a loss function is constructed using the mean square error, and this neural network is trained to obtain the runoff prediction. The average relative error is used as the error correction evaluation index, and its formula is: In the formula, is the total number of time steps for calculating the error; is the observed runoff (m³ / s); is the runoff prediction after error correction within the time period.

[0038] During the calibration period, is 6.4%. During the verification period, is 7.2%. During the test period is 4.8%, indicating that the runoff prediction method that couples the hydrological system response and the reservoir scheduling alternative network has a good simulation effect on the runoff in the Jin'anqiao Basin.

[0039] This embodiment provides a runoff prediction device, including a processor and a memory; a program or instruction is stored in the memory, and the program or instruction is loaded and executed by the processor to implement the steps of the runoff prediction method in the above embodiment.

[0040] This embodiment provides a computer-readable storage medium, and a program or instruction is stored on the readable storage medium. When the program or instruction is executed by a processor, the steps of the runoff prediction method in the above embodiment are implemented.

[0041] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Without departing from the concept of the present invention, those skilled in the art make several improvements and optimizations, which should all be regarded as the scope of protection of the present invention.

Claims

1. A runoff prediction method, characterized in that, Including the following steps: Step S1: Using meteorological inputs to construct a dynamic system response curve, and calculating the dynamic response characteristics of each grid cell at each moment, where the meteorological inputs include precipitation and temperature; Step S2: Extracting the static influence characteristics of runoff errors from the static attributes of the underlying surface data of the basin through convolution and non-linear transformation; Step S3: Using a neural network to align the inputs of the dynamic response characteristics and the static influence characteristics, and extracting the characteristics that reflect both temporal and spatial influences to obtain an error correction amount generated based on meteorological inputs and underlying surface information; Step S4: Combining the operation status and scheduling requirements of the reservoir to construct a reservoir scheduling alternative network, and outputting the correction component of the reservoir scheduling on runoff; Step S5: Collecting runoff forecast and measured data to construct a loss function, superimposing the above correction component and the output of the preliminary hydrological model, and simultaneously updating the neural network model parameters through end-to-end backpropagation to obtain a runoff prediction amount after error correction; 2. The runoff prediction method according to claim 1, wherein Step S3 includes: Step S301, use a neural network to align the dynamic features and static feature inputs: In the formula, is the concatenated feature vector after alignment and concatenation, , indicates that the concatenated feature map is a three-dimensional tensor with a spatial size of and a number of channels of ; is the input channel index, that is, the total number after concatenating the dynamic and static feature inputs; indicates concatenating the feature data in the channel dimension; and are the numbers of channels of the dynamic feature input data and the static feature input data respectively; is the longitudinal number of the grid, is the transverse number of the grid; is the grid at the time period dynamic response feature; is the basin static feature of the grid ; Step S302: Use convolution to fuse dynamic and static features to obtain features that can simultaneously reflect the influence of time and space: In the formula, is the output feature map; is the index of the output channel; is the activation function; is the learnable weight of the 1×1 convolution kernel in the channel dimension; is the bias of the convolution kernel; Step S303: Defining spatio-temporal correction weights and calculating the error correction amount generated based on meteorological inputs and underlying surface information; Among them, the spatio-temporal correction weight is defined as: In the formula, is the spatio-temporal correction weight, which reflects the spatial heterogeneity involved in the hydrological process within the basin through the catchment area of each grid cell and the propagation and contribution of the confluence process to errors; is the catchment area corresponding to this grid point; Based on the defined spatio-temporal correction weights , calculate the error correction component generated by meteorological input and underlying surface information: In the formula, is the error correction component generated by meteorological input and underlying surface information.

3. The runoff prediction method according to claim 2, wherein Dynamic response characteristics are as follows: In the formula, is the considered lag order, that is, in the time period convolution calculation is performed on the data of the past time periods; and are the precipitation and temperature of grid when the lag days are respectively; and are the weight coefficients of precipitation and temperature data at the grid scale respectively.

4. The runoff prediction method according to claim 2, wherein Static characteristics of the basin are as follows: In the formula, is the static attribute of grid , including soil type, vegetation type, and catchment area; represents a trainable neural network for outputting the encoded composite attribute vector of soil-vegetation-catchment area.

5. The runoff prediction method according to claim 2, characterized in that Step S4 includes: Step S401: Construct a water balance equation to quantify the reservoir operation status and scheduling requirements: In the formula, is the inflow during time period ; is the outflow during time period ; is the reservoir storage during time period , obtained through the water level - storage relationship; is the reservoir water level at time period , is the functional relationship between the water level and the storage, used to calculate the reservoir storage corresponding to this water level; and are respectively the minimum and maximum values of the outflow ; and are the minimum and maximum values of the reservoir water level ; Step S402: Combine the reservoir operation status and scheduling requirements to construct a reservoir scheduling alternative network, and output the correction component of the reservoir scheduling for runoff: In the formula, is a reservoir scheduling alternative model constructed based on a neural network, are trainable parameters; is the correction component of the reservoir scheduling for runoff.

6. The runoff prediction method according to claim 5, wherein Limit the outbound range through the activation function: In the formula, and are intermediate hidden layer variables, representing the output results of the first and second layers of the reservoir substitution network respectively; are the weights of the first layer, second layer and output layer of the reservoir substitution network respectively; is the bias of the corresponding reservoir substitution network; is the activation function of the reservoir substitution network.

7. The runoff prediction method according to claim 5, characterized in that Step S5 includes: Step S501: Collecting runoff forecast and measured data to construct a loss function; Step S502: Superimpose the system response correction component and the reservoir operation correction component on the output of the preliminary hydrological model, and update the neural network model parameters simultaneously through end-to-end backpropagation to obtain the runoff prediction: In the formula, is the runoff prediction after error correction within the time period; is the time period The simulated flow of the original hydrological model below; is the time period The error correction amount generated based on meteorological input and underlying surface information below; is the time period The correction amount of the impact of reservoir operation on runoff below.

8. The runoff prediction method according to claim 7, wherein, In step S501, the constructed loss function is as follows: In the formula, is the loss function; is the total number of time steps for calculating the error; is the observed runoff; is the runoff prediction value after error correction within the time period.

9. A runoff prediction device, characterized in that, Including a processor and a memory; the memory stores programs or instructions, and the programs or instructions are loaded and executed by the processor to implement the steps of the runoff prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The program or instruction is stored on the readable storage medium, and when the program or instruction is executed by the processor, it implements the steps of the runoff prediction method according to any one of claims 1 to 8.

Images