METHOD FOR CALIBRATING PARAMETERS OF HYDROLOGY FORECASTING MODEL BASED ON DEEP REINFORCEMENT LEARNING

Abstract

A method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning includes selecting according to basin characteristics, and the parameters and parameter value ranges to be calibrated by the model are determined. The method includes a calibrating the parameters of the hydrology forecasting model is established, and three elements of the reinforcement learning, that is, a state space, an action space and a reward function are determined. The method includes a deep reinforcement learning method DQN is applied to optimize the parameters to be calibrated by the hydrology forecasting model. In the present disclosure, by setting a stride length of the action value for the deep reinforcement learning model, an accuracy finally optimized by the calibration parameters can be freely controlled, and a DQN algorithm is adopted to search the entire space for the calibration parameters to ensure the optimality for optimizing the calibrated parameters.

Description

TECHNICAL FIELD

The present disclosure belongs to the technical field of calibrating parameters for hydrology forecasting models, and in particular to a method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning.

BACKGROUND

Hydrology forecasting models are widely applied in the fields, such as rainfall simulation forecasting, flood forecasting and early warning, hydrology process analysis, and play an important role in improving the efficiency of hydrology research. Since the structures of hydrology forecasting models are generally complex, the determination of the model parameters after the model is established has become a core problem to be solved urgently. Parameter calibration is to find a group of optimal parameter solutions in the hydrology forecasting model, so that the simulated and forecasted results are as close as possible to the measured data. For hydrology forecasting models, the accuracy of forecasting is directly affected by parameter calibration. Therefore, it is of great scientific significance and application value to study how to improve the speed and accuracy of parameter calibration.

The traditional manual trial and error method, gradient descent method, etc. are mostly adopted in the early methods for calibrating parameters of the hydrology forecasting models. Although these methods are intuitive, they require high experience of the staff and are easily affected by personal subjectivity, and the efficiency and accuracy for calibrating parameters are relatively low. With the development of computer technology, modern intelligent algorithms such as genetic algorithm and particle swarm algorithm are widely applied in the field of automatic parameter calibration of the hydrology forecasting models, which eliminates the deficiencies of the traditional methods. However, modern intelligent algorithms can retrieve a wide range of solutions, but there are problems of premature and being prone to falling into the local optimal solutions, which affects the selection of the global optimal solutions.

SUMMARY

The objectives of the present disclosure are to overcome the defects of the prior art and provide a method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning. In the present disclosure, the accuracy finally optimized by the calibration parameters can be freely controlled by setting the stride length of the action value for the deep reinforcement learning model, and a DQN algorithm is adopted to search the entire space for the calibration parameters so as to ensure optimality for optimizing the calibration parameters, so as to avoid the problem that the modern intelligent algorithms are premature and prone to falling into the local optimal solutions.

In order to solve the above technical problems, the following technical solutions are adopted in the present disclosure.

Provided is a method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning. The method includes the following steps.

In Step 1, a hydrology forecasting model is selected and parameters that need to be calibrated are determined; the hydrology forecasting model takes a rainfall time sequence and an evaporation time sequence as inputs, and takes a time sequence of a forecasted flow as an output.

In Step 2, a reinforcement learning model for calibrating the parameters of the hydrology forecasting model is established; the reinforcement learning refers to a process of an interactive learning between an intelligent agent Agent and environment, and three key elements for the reinforcement learning are a state space, an action space and a reward value function.

In Step 3, a deep reinforcement learning method DQN is applied to optimize the parameters to be calibrated by the hydrology forecasting model.

Furthermore, a process of selecting a hydrology forecasting model and determining the plurality of parameters to be calibrated in Step 1 includes the followings.

A different hydrology forecasting model is selected according to basin characteristics, and the parameters that need to be calibrated by the hydrology forecasting model are determined, where N is a number of parameters that need to be calibrated by the hydrology forecasting model;

The values for each parameter are in a range of:

$w_{\min }^i\le w^i\le w_{\max }^i,i=1,2,\dots ,N,$

• • where w_minⁱ and w_maxⁱ are a minimum value and a maximum value for an i-th parameter to be calibrated, respectively.

Moreover, the process of establishing the reinforcement learning model for calibrating the parameters of the hydrology forecasting model in Step S2 includes the following steps.

2-1) Determining the State Space for the Reinforcement Learning:

A state value for the reinforcement learning at a time t is defined as a one-dimensional vector s_t composed of the plurality of parameters to be calibrated in the hydrology forecasting model;

s _t=(w_t¹,w_t², . . . ,w_t^N)

• • where w_tⁱ, i=1, 2, . . . , N are values for the parameters to be calibrated by the hydrology forecasting model at the current time t.

Changes for the values for the parameters w_tⁱ at the time t have two possibilities: increase or decrease; when a magnitude of increasing or decreasing the parameters w_tⁱ is Δⁱ, and then a value w_t+1ⁱ for parameters at a time t+1 may be w_tⁱ+Δⁱ or w_tⁱ−Δⁱ.

2-2) Determining the Action Space for the Reinforcement Learning:

The action space A for the reinforcement learning is defined as all possibilities for changes of each of the parameters to be calibrated:

$A=\left[\begin{array}{cccccc}{\Delta }_1^1& {\Delta }_1^2& {\Delta }_1^3& \dots & {\Delta }_1^{N-1}& {\Delta }_1^N\\ -{\Delta }_2^1& {\Delta }_2^2& {\Delta }_2^3& \dots & {\Delta }_2^{N-1}& {\Delta }_2^N\\ -{\Delta }_3^1& -{\Delta }_3^2& {\Delta }_3^3& \dots & {\Delta }_3^{N-1}& {\Delta }_3^N\\ ⋮& ⋮& ⋮& \dots & ⋮& ⋮\\ {\Delta }_{2^N-1}^1& {\Delta }_{2^N-1}^2& {\Delta }_{2^N-1}^3& \dots & {\Delta }_{2^N-1}^{N-1}& {\Delta }_{2^N-1}^N\\ {\Delta }_{2^N}^1& -{\Delta }_{2^N}^2& {\Delta }_{2^N}^3& \dots & {\Delta }_{2^N}^{N-1}& {\Delta }_{2^N}^N\end{array}\right],$

where 2^N is a number of actions in the action space for the reinforcement learning; each row of matrix A is a selected action, that is, the possible values for an action value a_t at the time t.

2-3), Determining the Reward Value Function for the Reinforcement Learning:

Q_r=[q_r¹, q_r², q_r³, . . . , q_rⁱ, . . . , q_r^M] is set as flow values actually measured during M time periods, where q_rⁱ, i=1, 2, 3, . . . , M is a flow value actually measured during an i-th time period;

The predicted flow sequences are acquired through the hydrology forecasting model according to a state s_t=(w_t¹, w_t², . . . , w_t^N) at the time t and a state s_t+1=w_t+1¹, w_t+1¹, w_t+1¹, . . . , w_t+1^N) at the time t+1 respectively as

Q _t =[q _t ¹ ,q _t ² ,q _t ³ , . . . ,q _t ⁱ , . . . ,q _t ^M] and

Q _t+1 =[q _t+1 ¹ ,q _t+1 ² ,q _t+1 ³ , . . . ,q _t+1 ⁱ , . . . ,q _t+1 ^M],

where q_tⁱ, q_t+1ⁱi=1, 2, 3, . . . , M are flow values forecasted at the time t and the time t+1 during the i-th time period, respectively.

A root mean square error at the time t is defined as RMS_t, and a root mean square error at the time t+1 is defined as RMS_t+1, where

${\mathrm{RMS}}_t=\sqrt{\frac{\sum _{i=1}^M{\left(q_t^i-q_r^i\right)}^2}{M}}\mathrm{and}$ ${\mathrm{RMS}}_{t+1}=\sqrt{\frac{\sum _{i=1}^M{\left(q_{t+1}^i-q_r^i\right)}^2}{M}}.$

A reward value r_t(s_t, a_t, s_t+1) obtained by executing an action a_t in a state s_t at the time t to reach a state s_t+1 at the time t+1 is defined as follows:

$r_t\left(s_t,a_t,s_{t+1}\right)=\{\begin{array}{cc}{C}_1& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\\ 0& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\\ C_2-\sigma C_3& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\end{array}\mathrm{and}$ $\sigma =\{\begin{array}{cc}0& \forall w_t^i\in \left(w_{\min }^i,w_{\max }^i\right)\\ 1& \exists w_t^i\notin \left(w_{\min }^i,w_{\max }^i\right)\end{array}\begin{array}{c}\\ i=1,2,3,\dots ,M\\ \end{array},$

where C₁ is a constant greater than 0; C₂ is a constant less than 0; and 3 is a constant greater than 0.

Furthermore, the process of applying a deep reinforcement learning method DQN to optimize the parameters to be calibrated by the hydrology forecasting model in Step 3 includes the followings.

The following DQN algorithm for the reinforcement learning is performed according to the determined key elements of the reinforcement learning:

Input: a capacity of an experience pool D is initialized as N;

• • an action value network Q is initialized with a random weight θ; and
  • a target action value network {circumflex over (Q)} is initialized with a weight θ′=θ;
  • the process is as follows:
  • For 1, M do,
    • initializing a state s_t;
    • For t=1, T do,
      • selecting a random action a_t according to a probability ε, selecting a_t=argmax_aQ(s_t, a; θ) according to a probability 1−ε;
      • performing the action a_t and acquiring a reward r_t to obtain a next state s_t+1;
      • storing (s_t, a_t, r_t, s_t+1) in the experience pool D;
      • randomly selecting, when a certain number is reached in the experience pool D, a plurality of groups (s_t, a_t, r_t, s_t+1) as mini-batches;
  • it is set as follows:

$y_j=\{\begin{array}{c}{r}_j\\ r_j+\gamma {\max }_{a^{\prime }}\hat{Q}\left(s_{j+1},a^{\prime };{\theta }^{\prime }\right)\end{array},$

• • if a curtain ends in a j+1 i-th step, else;
    • performing a gradient descent step of (y_j−Q(s_j, a_j; θ))² to update the parameter θ of the action value network;
    • resetting the target action value network {circumflex over (Q)}=Q per C steps;
  • End For
  • End For;
  • Output: an optimal state of s_*=(w_*¹, w_*², w_*³, . . . , w_*^N).

Furthermore, when an optimization curve of the DQN algorithm reaches a state of convergence, a reward value in total is in a state of slight jitter at this time, and parameter values optimally calibrated are s_*=(w_*¹, w_*², w_*³, . . . , w_*^N).

Compared with the prior art, the present disclosure has the following advantages and beneficial effects.

1. In the present disclosure, by setting a stride length of the action value for the deep reinforcement learning model, an accuracy optimized by the calibration parameters can be freely controlled with respect to the calibration parameters for different properties to ensure the accuracy and rationality for optimizing the calibration parameters, and to avoid investing too much calculation resources into non-important parameters.

2. In the present disclosure, a DQN algorithm is adopted to search the entire space for the calibration parameters to ensure the optimality for optimizing the calibration parameters through self-decision-making and corrective abilities of the reinforcement learning, so as to avoid the problem that modern intelligent algorithms are premature and prone to falling into local optimal solutions.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a flowchart of a method according to an embodiment of the present disclosure.

FIG. 2 illustrates a schematic diagram of a hydrology forecasting model according to an embodiment of the present disclosure.

FIG. 3 illustrates a schematic diagram of a reinforcement learning according to an embodiment of the present disclosure.

FIG. 4 illustrates a schematic diagram of an action value network and a target action value network according to an embodiment of the present disclosure.

FIG. 5 illustrates a flowchart of an optimization algorithm of DQN according to an embodiment of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The present disclosure provides a method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning. The method includes the following steps. An appropriate hydrology forecasting model is selected according to basin characteristics, and the parameters and parameter value ranges calibrated by the model are determined; three elements of the reinforcement learning, that is, a state space, an action space and a reward function are constructed; a deep reinforcement learning method DQN is applied to optimize the parameters calibrated by the hydrology forecasting model. In the present disclosure, by setting a stride length of the action value for the deep reinforcement learning model, an accuracy finally optimized by the calibration parameters can be freely controlled, and a DQN algorithm is adopted to search the entire space for the calibration parameters to ensure the optimality for optimizing the calibrated parameters, so as to avoid the problem that modern intelligent algorithms are premature and prone to falling into local optimal solutions.

The present disclosure will be further described in detail below in combination with the accompanying drawings.

FIG. 1 illustrates a flowchart of a method according to an embodiment of the present disclosure. As illustrated in FIG. 1, the method in this embodiment includes the following steps.

In Step 1, a hydrology forecasting model is selected and parameters that need to be calibrated are determined.

As illustrated in FIG. 2, a different hydrology forecasting model is selected according to basin characteristics, the hydrology forecasting model takes a rainfall time sequence and an evaporation time sequence as inputs of the hydrology forecasting model, and takes a time sequence of a forecasted flow as an output of the hydrology forecasting model. The parameters wⁱ, i=1, 2, . . . , N that need to be calibrated by the hydrology forecasting model are determined, where N is a number of parameters that need to be calibrated by the hydrology forecasting model.

The value for each of the parameters is in a range of

$w_{\min }^i\le w^i\le w_{\max }^i,i=1,2,\dots ,N,$

where w_minⁱ and w_maxⁱ are a minimum value and a maximum value for an i-th parameter to be calibrated, respectively.

In Step 2, a reinforcement learning model for calibrating the parameters of the hydrology forecasting model is established.

As illustrated in FIG. 3, the reinforcement learning is the process of an interactive learning between an intelligent agent Agent and the environment. The agent can take corresponding actions according to the current state of the environment, so as to change the current state of the environment. Three key elements for the reinforcement learning are a state space, an action space and a reward value function.

(1) Determining the state space for the reinforcement learning:

The state value for the reinforcement learning at a time t is defined as a one-dimensional vector s_t composed of the plurality of parameters to be calibrated by the hydrology forecasting model;

s _t=(w_t¹,w_t², . . . ,w_t^N)

where w_tⁱ, i=1, 2, . . . , N are values for the parameters to be calibrated by the hydrology forecasting model at the current time t.

Changes for the values w_tⁱ for the parameters at the time t have two possibilities: increase or decrease; when a magnitude of increasing or decreasing the parameters w_tⁱ is both Δⁱ, then the values w_t+1ⁱ for the parameters at a time t+1 may be w_tⁱ+Δⁱ or w_tⁱ−Δⁱ).

(2) Determining the action space for the reinforcement learning is determined:

The action space A for the reinforcement learning is defined as all possibilities where each of the parameters to be calibrated changes:

$A=\left[\begin{array}{cccccc}{\Delta }_1^1& {\Delta }_1^2& {\Delta }_1^3& \dots & {\Delta }_1^{N-1}& {\Delta }_1^N\\ -{\Delta }_2^1& {\Delta }_2^2& {\Delta }_2^3& \dots & {\Delta }_2^{N-1}& {\Delta }_2^N\\ -{\Delta }_3^1& -{\Delta }_3^2& {\Delta }_3^3& \dots & {\Delta }_3^{N-1}& {\Delta }_3^N\\ ⋮& ⋮& ⋮& \cdots & ⋮& ⋮\\ {\Delta }_{2^N-1}^1& {\Delta }_{2^N-1}^2& {\Delta }_{2^N-1}^3& \cdots & {\Delta }_{2^N-1}^{N-1}& {\Delta }_{2^N-1}^N\\ {\Delta }_{2^N}^1& -{\Delta }_{2^N}^2& {\Delta }_{2^N}^3& \cdots & {\Delta }_{2^N}^{N-1}& {\Delta }_{2^N}^N\end{array}\right],$

where 2^N is a number of actions in the action space for the reinforcement learning; each row of matrix A is a selected action, that is, the possible values for an action value a_t at the time t. For example, when N=2 and Δ¹=Δ²=0.1, then

$A=\left[\begin{array}{cc}0.1& 0.1\\ -0.1& 0.1\\ 0.1& -0.1\\ -0.1& -0.1\end{array}\right].$

The possible values for a_t at this time are [0.1, 0.1], [−0.1, 0.1], [0.1, −0.1], [−0.1, −0.1].

(3) Determining the reward value function for the reinforcement learning:

Q_r=[q_r¹, q_r², q_r³, . . . , q_rⁱ, . . . , q_r^M] is set as flow values actually measured during M time periods, where q_rⁱ, i=1, 2, 3, . . . , M is a flow value actually measured during an i-th time period.

The predicted flow sequences acquired through the hydrology forecasting model according to a states s_t=(w_t¹, w_t², . . . , w_t^N) at the t time and a state s_t+1=(w_t+1¹, w_t+1², . . . , w_t+1^N) at the t+1 time, are respectively as

Q _t =[q _t ¹ ,q _t ² ,q _t ³ , . . . ,q _t ⁱ , . . . ,q _t ^M] and

Q _t =[q _t+1 ¹ ,q _t+1 ² ,q _t+1 ³ , . . . ,q _t+1 ⁱ , . . . ,q _t+1 ^M],

where q_tⁱ, q_t+1ⁱi=1, 2, 3, . . . , M are flow values forecasted at the time t and the time t+1 during the i-th time period, respectively.

A root mean square error at the time t is defined as RAI s_t, and a root mean square error at the time t+1 is defined as RMS_t+1 where

${\mathrm{RMS}}_t=\sqrt{\frac{\sum _{i=1}^M{\left(q_t^i-q_r^i\right)}^2}{M}}\mathrm{and}$ ${\mathrm{RMS}}_{t+1}=\sqrt{\frac{\sum _{i=1}^M{\left(q_{t+1}^i-q_r^i\right)}^2}{M}}.$

A reward value r_t(s_t, a_t, s_t+1) obtained by executing an action a_t in a state s_t at the time t to reach a state s_t+1 at the time t+1 is defined as follows:

$r_t\left(s_t,a_t,s_{t+1}\right)=\{\begin{array}{cc}{C}_1& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\\ 0& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\\ C_2-\sigma C_3& {\mathrm{RMS}}_t>{\mathrm{RMS}}_{t+1}\end{array}\mathrm{and}$ $\sigma =\{\begin{array}{cc}0& \forall w_t^i\in \left(w_{\min }^i,w_{\max }^i\right)\\ 1& \exists w_t^i\notin \left(w_{\min }^i,w_{\max }^i\right)\end{array}\begin{array}{c}\\ i=1,2,3,\dots ,M\\ \end{array},$

where C₁ is a constant greater than 0; C₂ is a constant less than 0; and C₃ is a constant greater than 0.

In Step 3, a deep reinforcement learning method DQN is applied to optimize the parameters to be calibrated by the hydrology forecasting model.

As illustrated in FIG. 4, provided is a schematic diagram of an action value network and a target action value network. In the network, the states are taken as the input neurons, and the input number is the number of parameters needed to be calibrated by the hydrology forecasting model. The action values are taken as the outputs, and the number of output values is the number of actions in the action space. The action value network is a value function used to evaluate a pair of the current state actions, and the reason why using a design of a neural network is that there are many states. The target action value network is used to gradually update the Q value. The parameters in the network are updated through the algorithm according to the updating formula of the loss function. After each C rounds of iterations, the parameters in the action value network are copied to the parameters in the target action value network. Q(s_j, a_j; θ) of the target action value network remains unchanged within a period of time, which reduces the possibility of oscillation and divergence of the loss value during training, thereby improving the stability of the algorithm.

As illustrated in FIG. 5, provided is a flowchart of a DQN optimization algorithm, in which MainNet is the action value network and targetNet is the target action value network. The DQN algorithm for the reinforcement learning is executed according to the key elements of the reinforcement learning determined in Step 2, and the procedure is as follows.

Input: a capacity of an experience pool D is initialized as N;

• • an action value network Q is initialized with a random weight θ, and the input and output of the network are illustrated by FIG. 4;
  • a target action value network {circumflex over (Q)} is initialized with a weight θ′=θ and the input and output of the network are illustrated by FIG. 4;
  • the process is as follows:
  • For 1, M do,
  • randomly initializing a state s₁=(w₀¹, w₀², w₀³, . . . , w₀^N)
  • For t=1, T do,
  • selecting a random action a_t according to a probability ε (ε is a smaller value), selecting a_t=argmax_aQ(s_t, a; θ) according to a probability 1−ε, which are calculated by the action value network;
  • performing the action a_t to obtain a next state s_t+1=(w_t+1¹, w_t+1², . . . , w_t+1^N; and acquiring a reward r_t where r_t is obtained from the calculation formula of Step 2;
  • storing (s_t, a_t, r_t, s_t+1) in the experience pool D, where at this time, it is necessary to determine whether the capacity of the experience pool is full, and when the capacity is full, the experience pool D can be updated by using the first-in and first-out strategy;
  • when a certain number is reached in the experience pool D, randomly selecting a plurality of groups (s_t, a_t, r_t, s_t+1) from the experience pool D as learning samples for a neural network;

It is obtained in the target action value network as:

$y_j=\{\begin{array}{c}{r}_j\\ r_j+\gamma {\max }_{a^{\prime }}\hat{Q}\left(s_{j+1},a^{\prime };{\theta }^{\prime }\right)\end{array}$

if a curtain ends in a j+1-th step, else;

• • performing a gradient descent step of (y_j−Q(s_j, a_j; θ))² by the action value network to update the network parameter θ;
  • resetting the target action value network {circumflex over (Q)}=Q per C steps, which means that the parameter θ for the action value network is set as the parameter θ′ for the target action value network;
  • End For
  • End For;
  • Output: an optimal state: s_*=(w_*¹, w_*², w_*³, . . . , w_*^N).

When an optimization curve of the algorithm of DQN reaches a state of convergence, a total reward value is in a state of slight jitter at this time, and parameter values optimally calibrated are s_*=(w_*¹, w_*², w_*³, . . . , w_*^N).

Claims

1. A method for calibrating parameters of a hydrology forecasting model based on a deep reinforcement learning, characterized by comprising following steps: Step 1, selecting a hydrology forecasting model and determining parameters that need to be calibrated, wherein the hydrology forecasting model takes a rainfall time sequence and an evaporation time sequence as inputs, and takes a time sequence of a forecasted flow as an output;Step 2, establishing a reinforcement learning model for calibrating the parameters of the hydrology forecasting model, wherein the reinforcement learning refers to a process of an interactive learning between an intelligent agent Agent and environment, and three key elements for the reinforcement learning are a state space, an action space and a reward value function; andStep 3, applying a deep reinforcement learning method DQN to optimize the parameters to be calibrated by the hydrology forecasting model.

2. The method for calibrating the parameters of the hydrology forecasting model based on the deep reinforcement learning according to claim 1, characterized in that the process of selecting the hydrology forecasting model and determining the plurality of parameters to be calibrated in Step 1 comprises: selecting a different hydrology forecasting model according to basin characteristics, and determining the parameters wi, i=1, 2, . . . , N that need to be calibrated by the hydrology forecasting model, where N is a number of the parameters that need to be calibrated by the hydrology forecasting model;wherein each of the parameters is in a range of:

3. The method for calibrating the parameters of the hydrology forecasting model based on the deep reinforcement learning according to claim 1, characterized in that the process of establishing the reinforcement learning model for calibrating the parameters of the hydrology forecasting model in Step 2 comprises: 2-1) determining the state space for the reinforcement learning:defining the state value for the reinforcement learning at a time t as a one-dimensional vector st composed of the plurality of parameters to be calibrated by the hydrology forecasting model; st=(wt1,wt2, . . . ,wtN)wherein wti, i=1, 2, . . . , N are values for the parameters to be calibrated by the hydrology forecasting model at the current time t; andchanges for the values wti for the parameters at the time t have two possibilities: increase or decrease; when a magnitude of increasing or decreasing the parameters wti is both Δi, a value wt+1i for the parameters at a time t+1 may be wti+Δi or wti−Δi;2-2) determining the action space for the reinforcement learning:defining the action space A for the reinforcement learning as all possibilities where each of the parameters to be calibrated changes:

4. The method for calibrating the parameters of the hydrology forecasting model based on the deep reinforcement learning according to claim 3, characterized in that the process of applying the deep reinforcement learning method DQN to optimize the parameters to be calibrated by the hydrology forecasting model in Step 3 comprises: performing, according to the determined key elements of the reinforcement learning, a following DQN algorithm for the reinforcement learning:inputting: initializing a capacity of an initialized experience pool D as N;initializing an action value network Q with a random weight θ;initializing a target action value network {circumflex over (Q)} with a weight θ′=θ;the process is as follows:For 1, M doinitializing a state s1;For t=1, T doselecting a random action at according to a probability ε, selecting at=argmaxaQ(st, a; θ) according to a probability 1−ε;performing the action at and acquiring a reward rt to get a next state st+1;storing (st, at, rt, st+1) in the experience pool D;randomly selecting, when a certain number is reached in the experience pool, a plurality of groups (st, at, rt, st+1) as mini-batches;setting

5. The method for calibrating the parameters of the hydrology forecasting model based on the deep reinforcement learning according to claim 4, characterized in that when an optimization curve of the DQN algorithm reaches a state of convergence, a reward value in total is in a state of slight jitter, and parameter values optimally calibrated are s*=(w*1, w*2, w*3, . . . , w*N).