Reservoir outbound flow prediction model construction method based on deep learning

Abstract

The invention discloses a reservoir outbound flow prediction model construction method based on deep learning. The method comprises the following steps: collecting reservoir hydrological characteristic data; according to the collected hydrological characteristic data of the reservoir, a long and short-term memory network model with an attention mechanism introduced is adopted, and an ex-reservoir flow prediction module is constructed; constructing a multi-factor intelligent prediction module by adopting a long-short-term memory network model introducing an attention mechanism according to the collected reservoir hydrological characteristic data; and based on a water balance principle, taking an output value of the reservoir outbound flow prediction module as an input reference quantity of the multi-factor intelligent prediction module, and constructing a reservoir outbound flow intelligent prediction model through a nested coupling mechanism. According to the invention, through dual-module nested coupling of the reservoir outbound flow prediction module and the multi-factor intelligent prediction module, the core problem that physical rule constraints and complex scheduling rules are difficult to be compatible in reservoir outbound flow prediction is innovatively solved.

Description

Technical Field

[0001] The present invention belongs to the field of reservoir scheduling and discloses a method for constructing a reservoir outflow prediction model based on deep learning. Background Art

[0002] Reservoirs are crucial engineering facilities for regulating water resource distribution, mitigating flood disasters, and ensuring agricultural irrigation and urban water supply. Their operation and management are crucial for the efficient utilization of water resources in a river basin and for disaster prevention and mitigation. However, reservoir operation and management are influenced by natural environmental factors, local socioeconomic factors, project-specific factors, policies and regulations, and emergencies. These factors contribute to complex operation rules and are significantly influenced by human activities. Consequently, traditional outflow prediction models rely primarily on physics and empirical statistics, requiring a large amount of corresponding operation data. These models suffer from computational complexity and poor applicability, making them difficult to adapt to complex operation rules.

[0003] While recently developed deep learning models can reduce reliance on physical parameters, they still have inherent limitations. Typical recurrent neural networks suffer from low Nash efficiency when processing hydrological sequences exceeding 72 hours due to the vanishing gradient problem, making them ineffective in capturing the lag effect of flood peaks. Furthermore, data-driven models lack embedded constraints on the law of water conservation, leading to unexpected and sudden changes in water storage capacity in forecasts. In particular, in multi-factor coordinated forecasting scenarios, inadequate modeling of water level, flow, and meteorological factors leads to large average forecast errors during the flood-dry season.

[0004] The core bottleneck that existing technologies have yet to overcome lies in the disconnect between physical mechanism modeling and intelligent data analysis. Physical models struggle to integrate new, heterogeneous data sources like real-time weather radar and satellite remote sensing, while purely data-driven approaches cannot guarantee the conservation of mass in hydrological processes. This disconnect results in severely inadequate robustness in reservoir outflow forecasting when responding to complex scheduling scenarios such as sudden droughts and floods. Summary of the Invention

[0005] This invention innovatively addresses the core challenge of reconciling physical constraints with complex scheduling rules in reservoir outflow forecasting through the nested coupling of a dual-module outflow forecasting module and a multi-factor intelligent forecasting module. By employing a physical constraint module driven by time characteristics and inflow, and strictly adhering to the law of water balance, it constructs a forecast benchmark. This fundamentally overcomes the shortcomings of traditional pure data models, which often distort water storage capacity due to sudden changes, and the difficulty of empirical models in adapting to extreme scheduling scenarios.

[0006] The technical solution provided by the present invention is a method for constructing a reservoir outflow prediction model based on deep learning, comprising the following steps: Step 1: Collect reservoir hydrological characteristic data; Step 1.1: Collect reservoir hydrological characteristic data and preprocess missing data; Step 1.2: Extract time series characteristic data based on the timestamps of the reservoir hydrological characteristic data collected in step 1.1.

[0007] Furthermore, the reservoir hydrological characteristic data include reservoir water level, water storage capacity, inflow, outflow and water level-reservoir capacity curve.

[0008] Step 2: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct an outflow flow prediction module; Step 2.1: Based on the reservoir hydrological characteristic data collected in step 1, the first long short-term memory network model is constructed with the time characteristic data and the inflow flow as input features and the outflow flow as the output target; Step 2.2: Introduce the temporal attention mechanism to dynamically calculate the feature weights of each time step and generate a temporal context feature vector; Step 2.3: Optimize the prediction accuracy of the first long short-term memory network model through hyperparameter tuning and build an outbound flow prediction module.

[0009] Step 3: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct a multi-factor intelligent prediction module; Step 3.1: Based on the reservoir hydrological characteristic data collected in step 1, a second long short-term memory network model is constructed with time series characteristic data, reservoir water level, water storage capacity and inflow as input features; Step 3.2: Introduce an attention mechanism module into the second long short-term memory network model to generate feature vectors of key time steps; Step 3.3: Build a multi-factor intelligent prediction module to output the optimized value of outbound flow through hyperparameter optimization.

[0010] Furthermore, the hyperparameters include the number of units, input shape, regularization parameters, loss function, optimizer, and training parameters.

[0011] Furthermore, the outbound flow prediction module is constructed based on time series flow characteristics to generate outbound flow prediction values; The multi-factor intelligent prediction module integrates multi-source hydrological environmental characteristics to generate an optimized value for outflow flow; The outbound flow prediction module and the multi-factor intelligent prediction module are trained and deployed through independent deep learning frameworks, and their outputs are nested and coupled to form the final prediction results.

[0012] Step 4: Based on the water balance principle, the output value of the outflow prediction module is used as the input reference quantity of the multi-factor intelligent prediction module, and a reservoir outflow intelligent prediction model is constructed through a nested coupling mechanism.

[0013] Step 4.1: Input the time series feature data and the inbound flow data into the outbound flow prediction module to generate an initial outbound flow prediction value; Step 4.2: Based on the initial outflow forecast value obtained in step 4.1 and historical hydrological characteristic data, calculate the dynamic change of water storage capacity through the water balance equation; The calculation formula of the water balance equation is: ; Where, INQ i For the current moment, OTQ i The outbound flow at the current moment, INQ i-1 To store historical moments, OTQ i-1 is the outbound flow at the historical moment, Δ t is the time interval between the current moment and the historical moment, W i-1 The reservoir water storage capacity at the historical moment.

[0014] Step 4.3: Update the reservoir state parameters based on the dynamic change of water storage obtained in step 4.2, and obtain the current reservoir water storage and its corresponding water level through the corresponding relationship between water level and storage capacity curve; Step 4.4: Input the updated reservoir state parameters, time series characteristics and inflow flow into the multi-factor intelligent prediction module, and output the optimized outflow flow prediction value.

[0015] Furthermore, the reservoir status parameters include reservoir water level and water storage capacity.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention embeds the water balance principle into a deep learning framework through the innovative design of a dual-channel nested coupling architecture for the outflow prediction module and the multi-factor intelligent prediction module. The outflow prediction module constructs a prediction benchmark based on the inflow and time series characteristics to capture basic reservoir scheduling information; the multi-factor intelligent prediction module integrates multiple environmental factors such as water level and water storage capacity to dynamically capture the nonlinear impact of scheduling rules. The two modules achieve closed-loop interaction through the water balance processor, effectively resolving the contradiction between the sudden distortion of water storage capacity in pure data models and the difficulty of traditional physical models in adapting to complex scheduling rules.

[0017] (2) The feature weighting technology introduced in this invention uses an attention mechanism to give the model the ability to dynamically focus on key hydrological events. Under complex conditions such as flood peak evolution during the flood season and sudden changes from drought to flood, the system adaptively adjusts the weights of historical time series features to accurately capture long-term evolution patterns such as flood peak lag and dry season duration. This overcomes the feature attenuation defect of recurrent neural networks in long-sequence prediction and significantly improves the prediction reliability of extreme hydrological events.

[0018] (3) This invention integrates reservoir operating status parameters with external environmental factors in real time, enabling the model to dynamically adapt to multi-objective scheduling rules. In new scenarios such as ecological water replenishment and pollution control, the system inverts the reservoir storage and discharge capacity based on the water level-capacity function, supporting real-time optimization decisions for multiple scheduling strategies.

[0019] (4) The nested coupling architecture of the present invention provides reliable benchmark values ​​through modularization, significantly optimizing module complexity. Compared with traditional hydraulic models, the present invention achieves continuous rolling prediction of outflow flow while ensuring prediction accuracy, providing real-time support for flood control emergency dispatch. The lightweight deployment solution is compatible with edge computing devices, significantly broadening the application scenarios of the technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below with reference to the accompanying drawings and examples.

[0021] Figure 1 This is a flow chart of a method for constructing a reservoir outflow prediction model according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the module construction of the reservoir outflow prediction model construction method according to an embodiment of the present invention.

[0023] Figure 3 This is the simulation result of the outflow process of the Caojie Power Station reservoir in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] like Figure 1 and Figure 2 and Figure 3 As shown in Figure 2, the method for constructing a reservoir outflow prediction model based on deep learning specifically includes: Based on the collected data on measured reservoir water levels, water storage, inflows, outflows, and water level-capacity curves, the present invention constructs a reservoir outflow prediction module that considers time characteristics and inflows, and a multi-factor intelligent prediction module that considers time characteristics, water levels, storage capacity, and inflows. Based on the principle of water balance, the two modules are nested and coupled to construct an intelligent reservoir outflow prediction model.

[0025] The method for constructing a reservoir outflow prediction model based on deep learning includes the following steps: Step 1: Collect reservoir hydrological characteristic data, which include the measured water level, water storage capacity, inflow, outflow and water level-capacity curve of the reservoir; Step 1.1: Collect reservoir hydrological characteristic data and preprocess missing data; Step 1.2: Extract time series characteristic data based on the timestamps of the reservoir hydrological characteristic data collected in step 1.1.

[0026] Calculate the weekly data based on the year, month, and day information of the reservoir hydrological characteristic data. Arrange the time characteristics of the data series, including month, day, hour, and week.

[0027] Collect the water level, water storage capacity, inflow and outflow data of the reservoir within a certain period of time and correspond them with the time characteristics, perform data processing and inspection, eliminate missing data, and ensure the rationality of the data.

[0028] Collect reservoir water level-capacity curve information.

[0029] Step 2: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct an outflow flow prediction module; Step 2.1: Based on the reservoir hydrological characteristic data collected in step 1, the first long short-term memory network model is constructed with time characteristic data and inflow as input features and outflow as output target. The reservoir hydrological characteristic data is divided into a rate period and a validation period, and the model is trained in the rate period. Step 2.2: Introduce the temporal attention mechanism to dynamically calculate the feature weights of each time step and generate a temporal context feature vector; Step 2.3: Optimize the prediction accuracy of the first long short-term memory network model through hyperparameter tuning and build an outbound flow prediction module.

[0030] Adjust the parameters of the first long short-term memory network model, including the number of units, input shape, regularization parameter, loss function, optimizer, and training parameters, to achieve the best model simulation effect. In addition, introduce the attention mechanism module into the model, use the hidden state sequence generated by LSTM, calculate the attention score of each time step through the attention network, and then obtain the attention weight α through Softmax normalization. t The hidden state of each time step is weightedly summed with the weight to obtain the context vector, which serves as the key basis for the final output of the model and ultimately builds the outbound traffic prediction module.

[0031] Step 3: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct a multi-factor intelligent prediction module; Step 3.1: Based on the reservoir hydrological characteristic data collected in step 1, a second long short-term memory network model is constructed with time series characteristic data, reservoir water level, water storage and inflow as input features, and the data is divided into a rate period and a validation period, and the model is trained in the rate period; Step 3.2: Introduce the attention mechanism module into the second LSTM network model to generate the feature vector of the key time step; Step 3.3: Build a multi-factor intelligent prediction module to output the optimized value of outbound flow through hyperparameter optimization.

[0032] Adjust the parameters of the second long short-term memory network model, including the number of units, input shape, regularization parameter, loss function, optimizer, and training parameters, to achieve the best model simulation effect. In addition, introduce the attention mechanism module into the model, use the hidden state sequence generated by LSTM, calculate the attention score of each time step through the attention network, and then obtain the attention weight β through Softmax normalization. t The hidden state of each time step is weightedly summed with the weight to obtain the context vector, which serves as the key basis for the final output of the model and ultimately constructs a multi-factor intelligent prediction module.

[0033] The outbound flow prediction module is built based on time series flow characteristics and is used to generate outbound flow prediction values; The multi-factor intelligent prediction module integrates multi-source hydrological environmental characteristics to generate optimized outflow flow values; The outbound flow prediction module and the multi-factor intelligent prediction module are trained and deployed through independent deep learning frameworks, and their outputs are nested and coupled to form the final prediction results.

[0034] Step 4: Based on the water balance principle, the output value of the outflow prediction module is used as the input benchmark of the multi-factor intelligent prediction module, and a reservoir outflow intelligent prediction model is constructed through a nested coupling mechanism.

[0035] Step 4.1: Input the time series feature data and inbound flow data into the outbound flow prediction module to generate the initial outbound flow prediction value; Step 4.2: Based on the initial outflow forecast value obtained in step 4.1 and historical hydrological characteristic data, calculate the dynamic change of water storage capacity through the water balance equation; The calculation formula of the water balance equation is: ; Where, INQ i For the current moment, OTQ i The outbound flow at the current moment, INQ i-1 To store historical moments, OTQi-1 is the outbound flow at the historical moment, Δ t is the time interval between the current moment and the historical moment, W i-1 The reservoir water storage capacity at the historical moment.

[0036] Step 4.3: Update the reservoir state parameters based on the dynamic change of water storage obtained in step 4.2, and obtain the current reservoir water storage and its corresponding water level through the corresponding relationship between water level and storage capacity curve; Step 4.4: Input the updated reservoir state parameters, time series characteristics and inflow into the multi-factor intelligent prediction module, and output the optimized outflow prediction value.

Claims

1. A method for constructing a reservoir outflow prediction model based on deep learning, characterized in that: The following steps are involved: Step 1: Collect reservoir hydrological characteristic data; Step 2: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct an outflow flow prediction module; Step 3: Based on the reservoir hydrological characteristic data collected in step 1, a long short-term memory network model with an attention mechanism is used to construct a multi-factor intelligent prediction module; Step 4: Based on the water balance principle, the output value of the outflow prediction module is used as the input reference quantity of the multi-factor intelligent prediction module, and a reservoir outflow intelligent prediction model is constructed through a nested coupling mechanism.

2. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 1, characterized in that: The step 1 includes the following sub-steps: Step 1.1: Collect reservoir hydrological characteristic data and preprocess missing data; Step 1.2: Extract time series characteristic data based on the timestamps of the reservoir hydrological characteristic data collected in step 1.

1.

3. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 2, characterized in that: The reservoir hydrological characteristic data include reservoir water level, water storage capacity, inflow, outflow and water level-reservoir capacity curve.

4. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 3 is characterized in that: The step 2 includes the following sub-steps: Step 2.1: Based on the reservoir hydrological characteristic data collected in step 1, the first long short-term memory network model is constructed with the time characteristic data and the inflow flow as input features and the outflow flow as the output target; Step 2.2: Introduce the temporal attention mechanism to dynamically calculate the feature weights of each time step and generate a temporal context feature vector; Step 2.3: Optimize the prediction accuracy of the first long short-term memory network model through hyperparameter tuning and build an outbound flow prediction module.

5. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 4, characterized in that: Step 3 includes the following sub-steps: Step 3.1: Based on the reservoir hydrological characteristic data collected in step 1, a second long short-term memory network model is constructed with time series characteristic data, reservoir water level, water storage capacity and inflow as input features; Step 3.2: Introduce an attention mechanism module into the second long short-term memory network model to generate feature vectors of key time steps; Step 3.3: Build a multi-factor intelligent prediction module to output the optimized value of outbound flow through hyperparameter optimization.

6. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 5, characterized in that: The hyperparameters include the number of units, input shape, regularization parameters, loss function, optimizer, and training parameters.

7. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 6, characterized in that: The outbound flow prediction module is constructed based on time series flow characteristics and is used to generate outbound flow prediction values; The multi-factor intelligent prediction module integrates multi-source hydrological environmental characteristics to generate an optimized value for outflow flow; The outbound flow prediction module and the multi-factor intelligent prediction module are trained and deployed through independent deep learning frameworks, and their outputs are nested and coupled to form the final prediction results.

8. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 7, characterized in that: The step 4 includes the following sub-steps: Step 4.1: Input the time series feature data and the inbound flow data into the outbound flow prediction module to generate an initial outbound flow prediction value; Step 4.2: Based on the initial outflow forecast value obtained in step 4.1 and historical hydrological characteristic data, calculate the dynamic change of water storage capacity through the water balance equation; Step 4.3: Update the reservoir state parameters based on the dynamic change of water storage obtained in step 4.2, and obtain the current reservoir water storage and its corresponding water level through the corresponding relationship between water level and storage capacity curve; Step 4.4: Input the updated reservoir state parameters, time series characteristics and inflow flow into the multi-factor intelligent prediction module, and output the optimized outflow flow prediction value.

9. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 8, characterized in that: The reservoir status parameters include reservoir water level and water storage capacity.

10. The method for constructing a reservoir outflow prediction model based on deep learning according to claim 9, characterized in that: The calculation formula of the water balance equation is: ； Where, INQ i For the current moment, OTQ i The outbound flow at the current moment, INQ i-1 To store historical moments, OTQ i-1 is the outbound flow at the historical moment, Δ t is the time interval between the current moment and the historical moment, W i-1 The reservoir water storage capacity at the historical moment.

Images