A method for medium and long term runoff state prediction by coupling intelligent algorithm and hydrological model

Abstract

The application discloses a method for coupling intelligent algorithm and hydrological model for medium and long term runoff state prediction, first, a multi-granularity hydrological space-time object data scene model for medium and long term runoff prediction of a basin is constructed to realize organic fusion of massive, heterogeneous and multi-granularity data; secondly, an artificial intelligence prediction method of 'deep learning + reinforcement learning' is adopted to obtain runoff state prediction and basin state evaluation; and then, the hydrological model and the Monte Carlo tree search algorithm are coupled for runoff process deduction, including runoff state selection, expansion, evaluation and backtracking, and finally, the prediction result is obtained. The application can realize accurate prediction and process deduction of medium and long term runoff state, and is helpful to improve the state prediction effect and macroscopic research and judgment ability of the medium and long term runoff change process under complex environment.

Description

Technical Field

[0001] This invention relates to a medium- and long-term hydrological forecasting method in the field of hydrological big data technology, specifically a medium- and long-term runoff state prediction method that couples intelligent algorithms with hydrological models, belonging to the field of hydrological data processing technology. Background Technology

[0002] Traditional hydrological models are based on the physical mechanisms of hydrological elements. Model construction often requires rich underlying surface information and suffers from problems such as long computation time, complex parameter settings, and high costs. Data-driven models based on intelligent algorithms, especially deep learning algorithms, offer advantages in feature learning and identification. They do not require knowledge of the formation process of hydrological elements, possess strong nonlinear fitting capabilities, and can improve the prediction speed, accuracy, and generalization ability of medium- and long-term runoff in watersheds to a certain extent.

[0003] However, the aforementioned data-driven models are generally black-box structures, driven solely by data and neglecting the physical mechanisms of runoff change. Establishing a mapping between explanatory and response variables based on massive amounts of data often fails to consider the physical laws, processes, constraints, and boundaries of runoff change, leading to outliers in runoff state predictions and making it difficult to maintain consistency with the fundamental principles of runoff change. Therefore, how to couple hydrological models based on physical mechanisms with data-driven models based on intelligent algorithms has become a pressing issue in the field of medium- and long-term runoff forecasting. Summary of the Invention

[0004] Purpose of the invention: To address the problems and shortcomings of existing technologies, this invention provides a method for predicting medium- and long-term runoff status by coupling intelligent algorithms and hydrological models.

[0005] Technical Solution: A method for predicting medium- and long-term runoff states by coupling intelligent algorithms and hydrological models. First, a multi-granularity hydrological spatiotemporal object data scenario model for medium- and long-term runoff prediction in watersheds is constructed to achieve the organic integration of massive, heterogeneous, and multi-granularity data. Second, an artificial intelligence prediction method combining deep learning and reinforcement learning is employed to obtain runoff state predictions and watershed state estimates. Then, the hydrological model is coupled with a Monte Carlo tree search algorithm to extrapolate the runoff process, including runoff state selection, expansion, evaluation, and backtracking, ultimately obtaining the prediction results. The method includes the following steps:

[0006] S1-1 Constructs a multi-granularity hydrological spatiotemporal object data scenario model for medium- and long-term runoff prediction in watersheds, and utilizes the massive, heterogeneous, and multi-granularity data set formed by the multi-granularity hydrological spatiotemporal object data scenario model.

[0007] S1-2 employs an artificial intelligence prediction method combining deep learning and reinforcement learning to obtain runoff state prediction and watershed state estimation. The specific implementation process is as follows: based on the dataset of S1-1, runoff state prediction is obtained through deep learning artificial intelligence prediction; based on the deep learning runoff state prediction, reinforcement learning artificial intelligence prediction is obtained; and based on the reinforcement learning runoff state prediction, watershed state estimation is constructed.

[0008] S1-3 couples the hydrological model with the Monte Carlo tree search algorithm to extrapolate runoff processes, including runoff state selection, expansion, evaluation, and backtracking, ultimately obtaining prediction results. The specific implementation process is as follows: a portion of the data in the S1-1 dataset is used as input to the hydrological model to obtain the hydrological model prediction results; the runoff state prediction from reinforcement learning, the hydrological model prediction results, and the watershed state estimate are coupled, and based on the coupling results, the runoff process is extrapolated using the Monte Carlo tree search algorithm.

[0009] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the medium- and long-term runoff state prediction method of the coupled intelligent algorithm and hydrological model as described above.

[0010] A computer-readable storage medium storing a computer program that performs a medium- to long-term runoff state prediction method that combines a coupled intelligent algorithm with a hydrological model as described above.

[0011] Beneficial effects: Compared with the prior art, the medium- and long-term runoff state prediction method coupled with intelligent algorithms and hydrological models provided by this invention has the following significant advantages: On the one hand, this invention can organically organize and integrate data with massive, heterogeneous, and multi-granular characteristics to solve the problems of spatiotemporal and semantic inconsistency. On the other hand, based on the physical mechanism of hydrological elements, by coupling intelligent algorithms and hydrological models, it can achieve accurate prediction and process extrapolation of medium- and long-term runoff state, which helps to improve the state prediction effect and macroscopic judgment ability of medium- and long-term runoff change process in complex environments. Attached Figure Description

[0012] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0014] like Figure 1 As shown, the method for predicting medium- and long-term runoff state by coupling intelligent algorithms and hydrological models includes the following:

[0015] Traditional data object processing methods mainly describe the characteristics of spatiotemporal objects from three aspects: time, space, and operation. However, they are difficult to describe the dynamic evolution process of spatiotemporal objects under different granularities.

[0016] The data object consistency processing method with multi-granularity spatiotemporal characteristics breaks through the description of static changes in spatiotemporal objects and realizes dynamic description of spatiotemporal objects. The original data consists of factors affecting runoff changes, including precipitation, temperature, topography, geology, soil, vegetation, and human activities. The above data forms a massive, heterogeneous, multi-granularity dataset through the construction of a multi-granularity hydrological spatiotemporal object data scenario model. The hydrological model mainly acquires model-related data related to runoff in the dataset, such as the underlying surface, watershed land use area, and soil area. The deep learning model takes the key feature factors affecting runoff state under given spatiotemporal granularity conditions in the dataset as input to the model, forming an x×y×n data matrix A, where x is the spatial granularity, which can be fine-grained or coarse-grained, with fine-grained corresponding to hydrological stations and coarse-grained corresponding to the watershed; y is the temporal granularity with a certain length, which can be selected from different time scales such as minutes, seconds, hours, days, months, and years; and n is the feature dimension formed by the spatiotemporal granularity features under the spatiotemporal granularity STG. To adjust the prediction accuracy for different runoff return periods and medium- to long-term runoff states, the spatial granularity (x), temporal granularity (y), and spatiotemporal granularity (STG) can be adjusted. Based on the changes in spatiotemporal granularity, a large data matrix A can be obtained for the watershed.

[0017] The multi-granularity hydrological spatiotemporal object data scenario model is defined as follows:

[0018]

[0019] In the formula:

[0020] OID—a unique spatiotemporal object identifier.

[0021] —A set of attributes that change with spatiotemporal granularity, consisting of a set of states. Composition, that is in: This indicates that within the effective time t, at a spatial granularity S g Set of spatial attributes under conditions k represents the number of all child objects of the spatiotemporal object O (spatial attribute); n represents the number of all different spatial granularities in the spatiotemporal object O (spatial attribute). i A corresponding set of spatial attributes and spatial granularity A set of binary relations.

[0022] S(t) — A collection of spatial data types representing the changes of spatiotemporal objects over time in a specific spatial coordinate system. Where: t i Indicates time, A collection of spatial data types representing spatiotemporal objects. q k q represents a sequence of discrete point coordinates. k ={(x1,y1,z1),(x2,y2,z2),…,(x n ,y n ,z n )}, where (x,y,z) represent the three-dimensional spatial coordinate axes.

[0023] A(t) — The set of attribute states of a spatiotemporal object O that change over time. in: Represents the spacetime object O in t i The state of the j-th attribute at time j, where j∈[1,m] and i∈[1,n].

[0024] T(T s ,T e — The temporal characteristic of the change in the state of a spatiotemporal object O. Where: T s T represents the initial time of a change in the state of a spatiotemporal object. e Indicates the end time of the state change of a spatiotemporal object.

[0025] Second, based on the massive, heterogeneous, and multi-granular data set formed by the multi-granularity hydrological spatiotemporal object data scenario model constructed above, an artificial intelligence prediction method of "deep learning + reinforcement learning" is adopted to obtain runoff state prediction and watershed state estimation. It mainly includes three parts: runoff state prediction based on supervised learning, runoff state prediction based on reinforcement learning, and construction of watershed state estimation.

[0026] Supervised learning-based runoff state prediction: Based on the aforementioned massive, heterogeneous, and multi-granular dataset, a deep learning model is trained using supervised learning. p(a|s) represents the probability distribution of the model's prediction for the next step, where a represents the most likely prediction result for all nodes, and s represents the current runoff state, related to matrix A. The model randomly selects (s, a) and uses stochastic gradient ascent to maximize the likelihood of a given runoff state s and runoff prediction a, thus simulating the natural runoff change process.

[0027] Runoff state prediction based on reinforcement learning: This method uses the runoff state in round t and the previously trained runoff state based on supervised learning for prediction. The DQN-network reinforcement learning algorithm is then used to refine the parameters of the runoff prediction in round t, ultimately yielding an enhanced runoff state prediction p. p (a|s). Reinforcement learning-based runoff state prediction will make multiple predictions, and the number of prediction steps will continue to increase. By comparing the predicted values ​​of subsequent runoff state changes with the measured values, the prediction will be traced back to the previous runoff state, and the judgment of the previous runoff state will be updated, so that the subsequent predictions will get closer and closer to the measured values.

[0028] Construction of watershed state estimation: Following runoff state prediction based on supervised learning and reinforcement learning, a watershed state estimation structure is constructed to reflect the accuracy of the runoff state prediction. The watershed state estimation is achieved through the estimation function v. p (s) represents the desired outcome, i.e., the condition of the watershed outlet section z. t To make a prediction, v p (s)=E[z t |s t =s,a t ...a T ], obtain a single watershed state estimate v(s), representing the situation of this round of prediction, a t , ..., a T Let represent the predicted runoff state from cycle t to cycle T, and s represent the current runoff state. During training, the evaluation function v... p The input data of (s) is the state-result pair (s,z) of the reinforcement learning output, and stochastic gradient descent is used to minimize the predicted value v(s) and the corresponding result z.

[0029] The aforementioned runoff state learning is based on historical data and predicts medium- to long-term runoff changes, without addressing new runoff processes, which may lead to overfitting. Therefore, it is necessary to obtain more medium- to long-term runoff change processes based on reinforcement learning samples to create new (s, a) patterns. III. Coupled Hydrological Model with Monte Carlo Tree Search Algorithm for Runoff Process Estimation

[0030] After obtaining runoff state prediction results based on supervised learning and reinforcement learning, the massive, heterogeneous, and multi-granular dataset formed by the multi-granularity hydrological spatiotemporal object data scene model is used as the basis. The hydrological model is coupled with the Monte Carlo tree search algorithm to extrapolate the runoff process, including runoff state selection, expansion, evaluation, and backtracking, and finally obtain the prediction results.

[0031] Hydrological model calculation: Based on the structural principle of the SWAT hydrological model, relevant data such as watershed elevation, land use, soil, and meteorology are collected from the massive, heterogeneous, and multi-granular data set formed by the multi-granularity hydrological spatiotemporal object data scene model. A spatial and attribute database is constructed and used as input data for the hydrological model. Then, the SWAT-CUP software is used to complete the calibration and verification of the model, thereby obtaining the hydrological model prediction results, which are used to supplement the relatively weak consideration of physical characteristics in the overall method of this invention.

[0032] Medium- to long-term runoff state prediction based on Monte Carlo tree search algorithm: After obtaining two prediction results (runoff state prediction based on supervised learning and reinforcement learning, and hydrological model prediction) and watershed state estimation, considering the complexity, variability and certain randomness of medium- to long-term runoff changes, and taking into account the return period of medium- to long-term prediction, the concept of Monte Carlo tree search is applied to combine runoff state prediction, hydrological model prediction and watershed state estimation.

[0033] Each search step in the Monte Carlo tree search algorithm can be divided into: a. selection; b. expansion; c. evaluation; d. backtracking.

[0034] a. Selection: In each medium- to long-term runoff state prediction, a global search is performed using the Monte Carlo tree search algorithm. A path is obtained when a leaf node is reached after step L. The selection process for each medium- to long-term runoff state prediction step is as follows:

[0035] a t =arg max a (Q(s t ,a)+u(s t ,a)) (2)

[0036] in:

[0037] Q(s t a)——The mixed value of various prediction results when continuously predicting up to the Lth step, which is related to the watershed state estimation.

[0038] u(s t a) — Values ​​obtained by exploring new prediction ranges are proportional to the prior probability of runoff state prediction.

[0039] In the initial state, reinforcement learning yields low results and the estimated watershed state is small; therefore, the main focus is on u(s). t a) is the main component, and as the number of training iterations increases, u(s) t The proportion of Q(s) gradually decreases, and the selection of Monte Carlo trees is mainly based on Q(s). t ,a) is the main one.

[0040] b. Extension: Based on the selection, and combined with the results of the hydrological model, extend the watershed state estimate at each step.

[0041] c. Evaluation: Continuous prediction up to the Lth step, based on the runoff state estimation function v p (s), to assess and form an updated watershed state estimate v L (s).

[0042] d. Backtracking: Based on the newly acquired v L (s), v in step L-1 before update L-1 (s) is used to correct the parameters and prediction results of the Monte Carlo tree search algorithm.

[0043] Medium- and long-term runoff prediction based on the Monte Carlo tree search algorithm ensures the accuracy of predicting new medium- and long-term runoff changes under unpredictable conditions, and uses hydrophysical characteristics to train deep learning and reinforcement learning models. When a certain threshold is reached, the node can be considered the final prediction result.

[0044] Obviously, those skilled in the art should understand that the steps of the medium- and long-term runoff state prediction method of the coupled intelligent algorithm and hydrological model in the above embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, and thus stored in a storage device for execution by the computing device. In some cases, the steps shown or described can be performed in a different order than presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, the embodiments of the present invention are not limited to any particular hardware and software combination.

Claims

1. A method for predicting medium- to long-term runoff state by coupling intelligent algorithms and hydrological models, characterized in that, First, a multi-granularity hydrological spatiotemporal object data scenario model is constructed for medium- and long-term runoff prediction in watersheds, achieving the organic integration of massive, heterogeneous, and multi-granularity data. Second, an artificial intelligence prediction method combining deep learning and reinforcement learning is adopted to obtain runoff state predictions and watershed state estimates. Then, the hydrological model is coupled with the Monte Carlo tree search algorithm to extrapolate the runoff process. Specifically, the steps include the following: S1-1 Construct a multi-granularity hydrological spatiotemporal object data scenario model for medium- and long-term runoff prediction in watersheds, and utilize the massive, heterogeneous, and multi-granularity data set formed by the multi-granularity hydrological spatiotemporal object data scenario model; S1-2 employs an artificial intelligence prediction method combining "deep learning and reinforcement learning" to obtain runoff state prediction and watershed state estimation. The specific implementation process is as follows: based on the dataset of S1-1, runoff state prediction is obtained through deep learning artificial intelligence prediction method; based on the deep learning runoff state prediction, reinforcement learning artificial intelligence prediction method is used to obtain reinforcement learning runoff state prediction; and based on the reinforcement learning runoff state prediction, watershed state estimation is constructed. S1-3 Coupled hydrological model with Monte Carlo tree search algorithm to extrapolate runoff process, including runoff state selection, expansion, evaluation and backtracking, and finally obtain prediction results; The specific implementation process is as follows: a portion of the data in the S1-1 dataset is used as input to the hydrological model to obtain the hydrological model prediction results; the runoff state prediction from reinforcement learning, the hydrological model prediction results, and the watershed state estimation are coupled; based on the coupling results, the runoff process is deduced using the Monte Carlo tree search algorithm. Deep learning models take key feature factors affecting runoff state under given spatiotemporal granularity conditions in the dataset as input to the model, forming an x×y×n data matrix. Where x is the spatial granularity; y is the temporal granularity with a set length; and n is the spatiotemporal granularity. The feature dimension formed by the spatiotemporal granularity features below; Based on the massive, heterogeneous, and multi-granular data set formed by the constructed multi-granularity hydrological spatiotemporal object data scenario model, the artificial intelligence prediction method of "deep learning + reinforcement learning" is adopted to obtain runoff state prediction and watershed state estimation, including three parts: runoff state prediction based on supervised learning, runoff state prediction based on reinforcement learning, and construction of watershed state estimation. Supervised learning-based runoff state prediction: A deep learning model is trained using supervised learning on massive, heterogeneous, and multi-granular datasets. Let represent the probability distribution of the model's prediction for the next step, where This represents the most likely prediction result for all nodes. Represents the current runoff state, and is associated with the matrix. Related; model random selection And use stochastic gradient ascent to maximize a given runoff state and runoff forecasting The likelihood, that is, simulating the natural runoff variation process; Runoff state prediction based on reinforcement learning: This method uses the runoff state in round t and the runoff state trained through supervised learning to make predictions. The DQN-network reinforcement learning algorithm is then used to refine the parameters of the runoff prediction in round t, ultimately resulting in an enhanced runoff state prediction. ; Construction of watershed state estimation: Following runoff state prediction based on supervised learning and reinforcement learning, a watershed state estimation structure is constructed to reflect the accuracy of the runoff state prediction. The watershed state estimation is achieved through an estimation function. This represents the desired outcome, specifically the condition of the watershed outlet section. Make predictions. To obtain a single watershed state estimate This represents the situation predicted in this round. a t , ..., a T The value represents the predicted runoff state from cycle t to cycle T, where s represents the current runoff state; during training, the evaluation function... The input data is the state-result pair of the reinforcement learning output. Stochastic gradient descent is used to improve the predicted values. Corresponding results z minimize; In the runoff process extrapolation by coupling the hydrological model and the Monte Carlo tree search algorithm, after obtaining the runoff state prediction results based on supervised learning and reinforcement learning, the runoff process extrapolation is carried out by coupling the hydrological model and the Monte Carlo tree search algorithm based on the massive, heterogeneous, and multi-granular data set formed by the multi-granularity hydrological spatiotemporal object data scene model, and finally the prediction results are obtained. Hydrological model calculation: Based on the structural principle of the SWAT hydrological model, data on watershed elevation, land use, soil, and meteorology are collected from the massive, heterogeneous, and multi-granular dataset formed by the multi-granularity hydrological spatiotemporal object data scene model. Spatial and attribute databases are constructed and used as input data for the hydrological model. Then, the SWAT-CUP software is used to complete the calibration and verification of the model, thereby obtaining the hydrological model prediction results. Medium- and long-term runoff state prediction based on Monte Carlo tree search algorithm: By introducing the concept of Monte Carlo tree search, runoff state prediction, hydrological model prediction and watershed state estimation are combined. The Monte Carlo tree search algorithm consists of the following steps for each search: selection, expansion, evaluation, and backtracking. Selection: In each medium- to long-term runoff state prediction, a global search is performed using the Monte Carlo tree search algorithm, after which... When a leaf node is reached, a path is obtained; the medium- to long-term runoff state prediction for each step is selected as follows: （2） in: —Continuous predictions up to the [number]th The mixed value of various prediction results at each step is related to the estimated state of the watershed; —The values ​​obtained from exploring new prediction ranges are proportional to the prior probability of runoff state prediction; Extension: Based on the selection, and combined with the results of the hydrological model, the watershed state estimate at each step is extended; Assessment: Continuous prediction up to the [number]th Step, based on the runoff state estimation function The assessment forms an updated estimate of the watershed status. ; Backtracking: Based on newly acquired Before the update Step The parameters and prediction results of the Monte Carlo tree search algorithm were corrected.

2. The method for predicting medium- and long-term runoff state by coupling intelligent algorithms and hydrological models according to claim 1, characterized in that, The raw data consists of factors that influence runoff changes, including precipitation, temperature, topography, geology, soil, vegetation, and human activities. The raw data is used to construct a multi-granularity hydrological spatiotemporal object data scenario model to form a massive, heterogeneous, and multi-granularity data set. The hydrological model mainly acquires model-related data related to runoff from the data set, including underlying surface, watershed land use area, and soil area.

3. The method for predicting medium- and long-term runoff state by coupling intelligent algorithms and hydrological models according to claim 1, characterized in that, To achieve different prediction accuracies for runoff return periods and medium- to long-term runoff states, spatial granularity x, temporal granularity y, and spatiotemporal granularity are adjusted. Adjustments are made; based on changes in spatiotemporal particles, the watershed acquires a large data matrix. .

4. The method for predicting medium- and long-term runoff state by coupling intelligent algorithms and hydrological models according to claim 1, characterized in that, The multi-granularity hydrological spatiotemporal object data scenario model is defined as follows: （1） In the formula: —A unique identifier for a spatiotemporal object; —A set of attributes that change with spatiotemporal granularity, consisting of a set of states. Composition, that is ;in: Indicates the effective time Within, at the spatial granularity Set of spatial attributes under conditions ; ; Representing spacetime objects The number of all child objects; Representing spacetime objects The number of all different spatial granularities in the space; A corresponding set of spatial attributes and spatial granularity A set of binary relations; —A collection of spatial data types in a spatial coordinate system, where spatiotemporal objects change over time. ;in: t i Indicates time, A collection of spatial data types representing spatiotemporal objects. , Represents a sequence of discrete point coordinates. , Represents the coordinate axes in three-dimensional space; Spacetime Objects A set of attribute states that change over time. ;in: Representing spacetime objects exist Time of the first Each attribute state, and , ; Spacetime Objects The temporal characteristics of state change; where: Indicates the initial time of a change in the state of a spatiotemporal object; Indicates the end time of the state change of a spatiotemporal object.

5. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the medium- and long-term runoff state prediction method of the coupled intelligent algorithm and hydrological model as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that performs a medium- to long-term runoff state prediction method based on the coupled intelligent algorithm and hydrological model as described in any one of claims 1-4.

Images