A method and system for predicting monthly runoff coupled with physical mechanisms

Abstract

The application discloses a kind of coupling physical mechanism's monthly runoff prediction method and system, the method is inputed to monthly runoff prediction model with last month's hydro-meteorological data, and the runoff of current month's watershed export section is obtained;Monthly runoff prediction model includes surface runoff module, soil module, underground module and confluence module connected in turn, with elevation, previous precipitation, previous temperature, soil moisture, groundwater storage, inflow runoff and the like as forecast factor input monthly runoff prediction model, each module is with convolutional encoder as basic architecture, combined with encoder-decoder structure, the spatial correlation characteristics of elements such as precipitation-runoff are described.Further, water balance constraint is used as the penalty term of loss function, to increase the physical rule penalty term constraint except data fitting error for model, to improve the logical inference reliability of model.

Description

TECHNICAL FIELD

[0001] The application relates to the field of runoff prediction, in particular to a monthly runoff prediction method and system coupled with physical mechanisms. BACKGROUND

[0002] Affected by special natural geographical and climatic conditions, the temporal and spatial distribution of water resources in China is uneven, the contradiction between water supply and demand is prominent, and water and flood disasters occur frequently and are serious. Runoff is the most direct manifestation of water resources abundance and scarcity, and accurate and reliable runoff prediction is of great significance for alleviating the contradiction between water supply and demand and reducing the loss of water and flood disasters. However, the monthly runoff process involves multiple spheres such as the hydrosphere, the atmosphere, and the lithosphere, and has many influencing factors and complex mechanisms. The existing prediction methods are affected by the nonlinear characteristics of the hydrological process and the uncertainty of precipitation, and the prediction performance is not strong, which seriously lags behind the actual production requirements. SUMMARY

[0003] The purpose of the application is to provide a deep learning monthly runoff prediction method and system coupled with physical mechanisms by comprehensively applying physical structural frameworks and physical rules, to seek a balance between deep learning and physical mechanism constraints and prediction performance improvement, and to realize a monthly runoff prediction method based on convolutional encoder and physical mechanism guidance, so as to improve the prediction performance of the monthly runoff.

[0004] The technical scheme of the application is a monthly runoff prediction method coupled with physical mechanisms, which comprises the following steps: inputting the hydro-meteorological data of the last month into a monthly runoff prediction model to obtain the runoff at the outlet section of the basin in the current month.

[0005] The monthly runoff prediction model comprises a surface runoff module, a soil module, an underground module and a confluence module connected in sequence, and the surface runoff module, the soil module, the underground module and the confluence module are all convolutional encoder models.

[0006] The surface runoff module predicts the precipitation, evapotranspiration and ground runoff data in the current month according to the evapotranspiration, potential evapotranspiration, air temperature, precipitation and elevation data in the last month.

[0007] The soil module predicts the soil moisture change and interflow data in the current month according to the soil moisture change data in the last month.

[0008] The underground module predicts the change of groundwater storage and underground runoff data in the current month according to the change of soil moisture in the current month and the change of groundwater storage in the last month.

[0009] The confluence module predicts the runoff at the outlet section of the basin in the current month according to the inflow runoff data in the last month, the change of groundwater storage, underground runoff, ground runoff and interflow data in the current month.

[0010] Further, the loss function of the monthly runoff prediction model is constrained by a water balance relationship, which is:

[0011] ET + ASM + AGW + Q = P + IQ

[0012] Q = P + IQ - ET - ASM - AGW

[0013] wherein Q refers to the theoretical value of the runoff at the outlet section of the basin satisfying the water balance constraint, P, ET, ASM, AGW and IQ represent the precipitation, evapotranspiration, soil moisture change, groundwater storage change and inflow of the basin in the current month, respectively.

[0014] Further, the loss function of the monthly runoff prediction model is:

[0015] Loss = lMSE + (1 - l)MSE physics data

[0016]

[0017] wherein Loss is the loss function, l represents the weight coefficient of the penalty for violating the water balance, Q pred refers to the predicted value of the runoff at the outlet section of the basin, Q obs refers to the observed value of the runoff at the outlet section of the basin, MSE data represents the root mean square error between Q pred and Q obs , MSE physics represents the root mean square error between Q pred and the theoretical value of the runoff at the outlet section of the basin satisfying the water balance constraint Q, and n is the number of observed values.

[0018] Further, the architecture of the surface runoff module, soil module, groundwater module and confluence module is a convolutional encoder-convolutional decoder architecture

[0019] Further, the convolutional encoder-convolutional decoder architecture includes a convolutional encoder, a convolutional decoder and a fully connected layer; the convolutional encoder and the convolutional decoder each include two sequences of convolutional operations for increasing and decreasing the number of feature maps, respectively.

[0020] Further, the hydro-meteorological data of the previous month is preprocessed, including the evapotranspiration, potential evapotranspiration, air temperature, precipitation, elevation, soil moisture change and groundwater storage change data of the previous month; the preprocessing includes integrating the hydro-meteorological data of the previous month to the same spatial scale by spatial interpolation method and standardizing the data.

[0021] ​Further, the Z-Score standardization method is used for standardization.

[0022] The coupled physical mechanism monthly runoff prediction system comprises a monthly runoff prediction model; hydro-meteorological data of a previous month are input into the monthly runoff prediction model to obtain basin outlet runoff of a current month;

[0023] The monthly runoff prediction model comprises, in sequence, a surface runoff module, a soil module, an underground module and a confluence module, and the surface runoff module, the soil module, the underground module and the confluence module are all convolutional encoder models.

[0024] The surface runoff module predicts, according to evapotranspiration, potential evapotranspiration, air temperature, precipitation and elevation data of the previous month, precipitation, evapotranspiration and surface runoff data of the current month;

[0025] The soil module predicts, according to soil moisture change data of the previous month, soil moisture change and interflow data of the current month;

[0026] The underground module predicts, according to soil moisture change data of the current month and underground water storage change data of the previous month, underground water storage change and underground runoff data of the current month;

[0027] The confluence module predicts, according to inflow runoff data of the previous month, underground water storage change, underground runoff, surface runoff and interflow data of the current month, basin outlet section runoff of the current month.

[0028] The electronic device comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, and the computer program realizes the coupled physical mechanism monthly runoff prediction method when loaded into the processor.

[0029] The computer readable storage medium stores a computer program, and the computer program realizes the coupled physical mechanism monthly runoff prediction method when executed by the processor.

[0030] Advantages: Compared with the prior art, the advantages of the present application are that:

[0031] (1) The coupling physical mechanism is realized in the application, and physical mechanism guided deep learning monthly runoff prediction is realized. The physical reliability of the deep learning runoff prediction model is not strong, and the prediction result not conforming to the physical law is easy to be generated. In order to improve the quality of the deep learning monthly runoff prediction, the nonlinear mapping relationship between the monthly runoff and the land-air influence elements is constructed, and the water balance relationship between the hydrological cycle elements such as evapotranspiration and precipitation and the runoff is established. According to this, the loss function is improved, and the hydrological process physical architecture representing the hydrological response relationship is designed as the physical rule constraint, the model causal logic inference process is strengthened, so that the deep learning prediction model coupled with the physical mechanism is established, the physical reliability of the model is improved, and the prediction performance is considered.

[0032] (2) The spatial modeling of the basin hydrological process is realized by using the convolution network in the application. The traditional deep learning is mainly based on single point and single station prediction. The spatial correlation of the hydrological and meteorological elements such as precipitation and air temperature is considered by using the convolution network, the spatial structure relationship of rainfall-runoff is described, so that the spatial information characteristics of the hydrological and meteorological elements can be more fully utilized in the runoff prediction process, and the spatial relationship of the hydrological elements is reflected.

[0033] In summary, the physical mechanism and the convolution deep network are coupled in the application, the physical reliability of the deep learning model is improved, the prediction performance is considered, the application value of the deep learning in the hydrological prediction is improved, the hydrological process modeling at the basin scale is realized, and the application has rationality and effectiveness. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The monthly runoff prediction method flow chart of the application is shown in the figure;

[0035] Figure 2 The monthly runoff prediction model structure chart of the embodiment of the application is shown in the figure;

[0036] Figure 3 The runoff measured value and the prediction value comparison chart of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0037] In order to facilitate understanding of the application, the following description is made:

[0038] Deep learning:

[0039] Definition 1: Deep learning usually refers to deep neural network, which is a kind of model with multi-layer network structure developed on the basis of neural network, including long short-term memory (LSTM), convolutional neural network (CNN) and other common structures.

[0040] Convolutional encoder:

[0041] Definition 2: Convolutional Autoencoder (CAE) is a deep learning model that can automatically extract features of input data and compress them into a low-dimensional representation through the combination of an encoder and a decoder.

[0042] Water balance relationship:

[0043] Definition 3: Water balance refers to the difference between the income and expenditure of water in a region within any time period, which is equal to the change in water storage in the region during that period.

[0044] As shown in Figure 1 , the coupling physical mechanism monthly runoff prediction method includes the following steps:

[0045] (1) Forecast factor selection: Selecting previous precipitation, temperature, evapotranspiration, and other meteorological factors, as well as topography, soil moisture, groundwater storage, and other underlying surface factors as potential influencing factors of regional precipitation. Comprehensive comparison and selection of elements with physical significance as model prediction factors.

[0046] (2) Data preprocessing: Collect and preprocess the data based on the selected prediction factors. First, unify the spatial resolution of the data and integrate them to the same spatial scale based on spatial interpolation methods; second, standardize all data.

[0047] (3) Design the basic architecture of the model: Apply the module combination and superposition principle to represent the hydrological cycle processes of surface processes, interflow, groundwater, and confluence with four sub-modules. Finally, connect the outputs of each module to obtain the runoff prediction results.

[0048] (4) Design the architecture of the sub-module: Each sub-module uses convolutional autoencoder as the basic architecture and adopts encoding-decoding structure. The encoder includes multiple sequential convolutions, and after each convolution, the number of filters doubles in the convolution layer. The decoder is similar, including the same sequence of deconvolution, and after each deconvolution, the feature map is reduced by half. The last layer uses convolution operation to output the result.

[0049] (5) Modify the loss function: Add the water balance constraint as a penalty term of the loss function to increase the physical rule penalty term constraint for the model in addition to the data fitting error.

[0050] (6) Model training: Hydro-meteorological data is used to form a sample set. The samples are divided into three subsets: training, validation, and testing. The training set is used for model parameter fitting, the validation set is used to evaluate the prediction performance of the trained model to optimize the hyperparameters, and the test set is used to evaluate the fitting performance of the optimized model. The root mean square error, mean relative error, Nash efficiency coefficient, and other evaluation indicators are used to optimize the model parameters and select the hyperparameters using the Adam algorithm, regularization, early stopping, and other strategies.

[0051] According to the constructed model, the current measured precipitation, air temperature, and other data are applied to predict the runoff at the outlet section of the basin at the next time.

[0052] Taking the outlet runoff of the upper reaches of the Yangtze River as an example, the deep learning monthly runoff prediction method coupled with physical mechanisms is used to predict the monthly runoff at the outlet section (Cuntan Station) of the upper reaches of the Yangtze River.

[0053] 1. Measured data

[0054] Monthly precipitation and monthly air temperature observation data are obtained from the China Ground Precipitation and Temperature Monthly Value 0.5°×0.5° Grid Dataset (V2.0) published by the National Meteorological Information Center. Hydrological station observation runoff data are obtained from the Information Center of the Ministry of Water Resources. Potential evaporation, actual evaporation, soil moisture, and groundwater storage data are obtained from the GLDAS Noah Land Surface Model product. All hydro-meteorological data are from January 1961 to December 2016. 1 km digital elevation model data are obtained from the Resource and Environment Science and Data Center of the Chinese Academy of Sciences.

[0055] 2. Selection of prediction factors

[0056] The potential influencing factors of the model are elevation data, previous precipitation, temperature, potential evaporation, actual evaporation, soil moisture, and groundwater storage data, i.e., hydro-meteorological elements at t-1 time. The model output is the runoff at the outlet section of the basin at t time. Where t represents the current month, and t-1 represents the previous month.

[0057] 3. Data preprocessing

[0058] The data in step 1 above is re-scaled to a 0.5°×0.5° grid dataset through bilinear interpolation. The mean and variance of each type of data are calculated to complete data standardization.

[0059] In this embodiment, the Z-Score standardization method is used, as shown in the following formula:

[0060]

[0061] 4. Model basic architecture design

[0062] The basic architecture of the constructed model is as followsFigure 2 (a) shows. Among them, the model includes four modules representing the hydrological cycle processes of surface processes, interflow, groundwater, and confluence.

[0063] The surface runoff module takes the previous evapotranspiration ET(t-1), potential evapotranspiration PET(t-1), air temperature T(t-1), precipitation P(t-1), and elevation data as input, and the submodule output is the current time precipitation P(t), evapotranspiration ET(t), and surface runoff RD(t) (all as intermediate variables in the model);

[0064] The soil module takes the previous soil moisture change ΔSM(t-1) as input, and the submodule output is the current time soil moisture change ΔSM(t) and interflow RS(t) (both as intermediate variables in the model);

[0065] The groundwater module takes the current time soil moisture change ΔSM(t) output by the soil module and the previous groundwater storage change ΔGW(t-1) as input, and the submodule output is the current time groundwater storage change ΔGW(t) and groundwater runoff RG(t) (both as intermediate variables in the model);

[0066] The confluence module takes the previous inflow runoff IQ(t-1), the current time groundwater storage change ΔGW(t) and groundwater runoff RG(t) output by the groundwater module, the surface runoff RD(t) output by the surface module, and the interflow RS(t) output by the soil module as input, and finally outputs the current time runoff Q(t) at the outlet section of the watershed.

[0067] Among them, the current time evapotranspiration, precipitation, soil moisture change, groundwater storage change, etc. are added as loss function penalty terms to the model training.

[0068] 5、Submodule architecture design.

[0069] Each submodule takes a convolutional encoder as the basic architecture, as shown in Figure 2 (b). Among them, the encoder includes two sequences of convolution operations, and each sequence contains two 3x3 convolution kernels with a step size of 1. After each operation, the number of filters in the convolution layer is doubled (8->16). The decoder is similar, including two sequences of convolution operations, and each sequence performs a 3x3 convolution operation to reduce the feature map by half (16->8). Then use a 1x1 convolution operation and a fully connected layer to output the result. In the convolution operation, use ReLU for activation; in the last convolution layer and the fully connected layer, use the linear function for activation.

[0070] 6、Modify the loss function Loss.

[0071] The water balance relationship adopted by the model is represented as:

[0072] ET + ASM + AGW + Q = P + IQ

[0073] Q = P + IQ - ET - ASM - AGW

[0074] Where, Q refers to the theoretical value of runoff at the outlet section of the basin that meets the water balance constraint; P, ET, ASM, AGW and IQ represent precipitation, evapotranspiration, soil moisture change, change in groundwater storage, and inflow of the basin, respectively.

[0075] To meet the water balance constraint, the predicted runoff at the outlet section of the model (Q pre ) should be close to the theoretical value (Q), then the loss function of the model is as follows:

[0076] Loss = lMSE physics + (1 - l)MSE data

[0077]

[0078] Where, Q pred refers to the predicted value of runoff, Q obs refers to the observed value of runoff, MSE represents the root mean square error, MSE data represents the error between the predicted value of runoff (Q pred ) and the observed value of runoff (Q obs ); MSE physics represents the error between the predicted value of runoff (Q pred ) and the theoretical value of runoff that meets the water balance constraint (Q); n is the number of observed values; l represents the weight coefficient of the punishment for violating the water balance.

[0079] 7. Model training

[0080] The sample set is composed of monthly precipitation, temperature, etc. The learning rate is set to 1e-3, and the sample set is divided into training set, validation set and test set. The period from 1961 to 1992 is the model training period, the period from 1993 to 2000 is the validation period, and the period from 2001 to 2016 is the model test period.

[0081] 8. Model application and evaluation

[0082] The root mean square error, relative error, Nash coefficient and correlation coefficient are used to compare and evaluate the effect of the model. The specific results are shown in Table 1 and Figure 3 .

[0083] Table 1: Effect of the constructed model on monthly runoff prediction in the upper reaches of the Yangtze River Basin

[0084]

[0085] It can be seen from the above results that the application uses coupling physical mechanism and deep learning to predict monthly runoff, can consider physical rule constraints, and realizes accurate prediction of monthly runoff.

[0086] The monthly runoff prediction system coupled with a physical mechanism comprises a monthly runoff prediction model; hydro-meteorological data of the last month is input into the monthly runoff prediction model to obtain the runoff at the outlet of the basin in the current month;

[0087] The monthly runoff prediction model comprises a surface runoff module, a soil module, a groundwater module and a confluence module connected in sequence, and the surface runoff module, the soil module, the groundwater module and the confluence module are all convolutional encoder models;

[0088] The surface runoff module predicts precipitation, evapotranspiration and ground runoff data in the current month according to evapotranspiration, potential evapotranspiration, air temperature, precipitation and elevation data in the last month;

[0089] The soil module predicts soil moisture change and interflow data in the current month according to soil moisture change data in the last month;

[0090] The groundwater module predicts groundwater storage change and groundwater runoff data in the current month according to soil moisture change data in the current month and groundwater storage change data in the last month;

[0091] The confluence module predicts runoff at the outlet section of the basin in the current month according to inflow runoff data in the last month, groundwater storage change, groundwater runoff, ground runoff and interflow data in the current month.

[0092] The electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program realizes the monthly runoff prediction method coupled with a physical mechanism when loaded into the processor.

[0093] The computer readable storage medium stores a computer program, and the computer program realizes the monthly runoff prediction method coupled with a physical mechanism when executed by the processor.

[0094] The computer readable storage medium can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory or any other medium that can be used to store desired program codes in the form of instructions or data structures and can be accessed by a computer.

[0095] The processor is used to execute the computer program stored in the memory to realize each step in the method involved in the above embodiments.

Claims

1. A method for predicting monthly runoff coupled with physical mechanisms, characterized in that, The steps include: inputting the hydrological and meteorological data of the previous month into the monthly runoff prediction model to obtain the runoff at the watershed outlet section for the current month; The monthly runoff prediction model includes a surface runoff module, a soil module, a subsurface module, and a runoff module connected in sequence. The surface runoff module, soil module, subsurface module, and runoff module are all convolutional encoder models. The surface runoff module predicts precipitation, evapotranspiration, and surface runoff data for the current month based on the previous month's evapotranspiration, potential evapotranspiration, temperature, precipitation, and elevation data. Based on the soil moisture change data of the previous month, the soil module predicts the soil moisture change and interflow data for the current month. The underground module predicts the changes in groundwater storage and groundwater runoff for the current month based on the soil moisture change data for the current month and the groundwater storage change data for the previous month. The runoff module predicts the runoff at the watershed outlet section for the current month based on the inflow runoff data of the previous month, the changes in groundwater storage, groundwater runoff, surface runoff and interflow data for the current month. The loss function of the monthly runoff prediction model is constrained by the water balance relationship, which is as follows: ET + ΔSM + ΔGW + Q = P + IQ Q = P + IQ - ET - ΔSM - ΔGW Where Q refers to the theoretical value of runoff at the watershed outlet section that meets the water balance constraint in the current month, and P, ET, ΔSM, ΔGW and IQ represent the precipitation, evapotranspiration, soil moisture change, groundwater storage change and watershed inflow in the current month, respectively. The loss function of the monthly runoff prediction model is: Loss=λMSE physics +(1-λ)MSE data Where Loss is the loss function, λ represents the weighting coefficient of the penalty for violating water balance, and Q pred The predicted runoff value at the outlet section of the watershed, Q obs MSE refers to the observed runoff at the outlet section of the watershed. data Q represents pred With Q obs The root mean square error between them, MSE physics Q represents pred The root mean square error between the measured value and the theoretical value Q of the runoff at the watershed outlet section that satisfies the water balance constraint, where n is the number of observations.

2. The monthly runoff prediction method based on coupled physical mechanisms according to claim 1, characterized in that, The surface runoff module, soil module, underground module, and runoff module all have a convolutional encoder-convolutional decoder architecture.

3. The monthly runoff prediction method based on coupled physical mechanisms according to claim 2, characterized in that, The convolutional encoder-decoder architecture includes a convolutional encoder, a convolutional decoder, and a fully connected layer; both the convolutional encoder and the convolutional decoder include two sequences of convolutional operations, used to increase and decrease the number of feature maps, respectively.

4. The monthly runoff prediction method based on coupled physical mechanisms according to claim 1, characterized in that, The hydrological and meteorological data of the previous month are preprocessed. The hydrological and meteorological data of the previous month include data on evapotranspiration, potential evapotranspiration, temperature, precipitation, elevation, soil moisture changes and groundwater storage changes. The preprocessing includes integrating the hydrological and meteorological data of the previous month into the same spatial scale through spatial interpolation and standardizing the data.

5. The monthly runoff prediction method based on coupled physical mechanisms according to claim 4, characterized in that, Standardization was performed using the Z-Score standardization method.

6. A monthly runoff prediction system based on the coupled physical mechanism of the method described in claim 1, characterized in that, This includes a monthly runoff prediction model; the hydrological and meteorological data of the previous month are input into the monthly runoff prediction model to obtain the watershed outlet runoff for the current month; The monthly runoff prediction model includes a surface runoff module, a soil module, a subsurface module, and a runoff module connected in sequence. The surface runoff module, soil module, subsurface module, and runoff module are all convolutional encoder models. The surface runoff module predicts precipitation, evapotranspiration, and surface runoff data for the current month based on the previous month's evapotranspiration, potential evapotranspiration, temperature, precipitation, and elevation data. Based on the soil moisture change data of the previous month, the soil module predicts the soil moisture change and interflow data for the current month. The underground module predicts the changes in groundwater storage and groundwater runoff data for the current month based on the changes in soil moisture in the current month and the changes in groundwater storage in the previous month. The confluence module predicts the runoff at the watershed outlet section for the current month based on the inflow runoff data of the previous month, the changes in groundwater storage, groundwater runoff, surface runoff, and interflow data for the current month.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the monthly runoff prediction method based on the coupled physical mechanism according to any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the monthly runoff prediction method based on the coupled physical mechanism according to any one of claims 1-5.

Images