Reservoir flow control method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of reservoir automatic control, and discloses a reservoir flow control method and system based on artificial intelligence. The method comprises the following steps: collecting real-time multi-source monitoring data of a reservoir basin, and generating a corresponding real-time basin topological graph; according to the real-time watershed topological graph, a reservoir flow and water level prediction model is used for prediction, and a real-time reservoir flow and water level prediction result is obtained; if the real-time dynamic water level safety threshold value is exceeded, a reservoir flow control process is triggered; according to the real-time flow and the water level prediction result, a reservoir flow control model is used for strategy generation, and a real-time reservoir flow control strategy is obtained; and according to the real-time reservoir flow control strategy, a real-time reservoir flow control instruction is generated and executed. According to the method, the problems of limited prediction capability, response lag, control strategy stiffness, lack of self-adaptability and high safety risk in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of automatic control of reservoirs, and particularly relates to a reservoir flow control method and system based on artificial intelligence. BACKGROUND

[0002] Reservoirs play a crucial role in flood control, water supply, power generation, irrigation, etc. Effective control of reservoir flow is the key to ensuring the safe operation of reservoirs and maximizing their comprehensive benefits. Traditional reservoir flow control methods rely heavily on experience or simple models based on historical data, making it difficult to cope with complex and variable hydro-meteorological conditions in the basin.

[0003] With the development of artificial intelligence technology, especially the ability of deep learning and reinforcement learning in processing complex spatio-temporal data and decision optimization, new solutions have been provided for reservoir flow control. However, how to effectively integrate these advanced technologies into reservoir flow control and build a full-chain intelligent control system that can real-time perceive basin conditions, accurately predict, intelligently decide and automatically execute, remains a major challenge in the current technical field.

[0004] The prior art has the following defects: 1) Limited prediction ability and delayed response: Traditional methods often rely on engineers' experience or use models based on historical average data and simplified physical processes (such as empirical formulas and one-dimensional hydrological models). These methods are difficult to accurately capture the complex hydrological processes within the basin, especially in extreme rainfall, snowmelt or dam breach situations, where prediction accuracy drops sharply. Only a single or few monitoring points (such as upstream water level, rainfall) are considered, lacking effective integration and utilization of multi-source heterogeneous data (such as weather radar echoes, satellite remote sensing images, distributed rainfall stations, river flow, soil moisture, etc.) within the entire basin. This makes the prediction of the overall water regime and inflow process in the basin not comprehensive and accurate enough. From data collection, manual analysis, scheme development to operation, it often takes a long time. In emergency situations such as floods, this lag may result in missed opportunities for optimal flood release, exacerbating flood risk.

[0005] 2) Control strategy is rigid and lacks adaptability: Many reservoirs use rules based on fixed water levels or inflow to release water (such as "water level exceeds X meters, open Y gate to release X cubic meters per second"). This strategy lacks flexibility and cannot dynamically adjust to real-time water regime and meteorological information from multiple points in the basin. In complex conditions such as multiple heavy rains in different areas of the basin, superimposed inflow in different regions, limited river flood carrying capacity, etc., simple rules cannot develop optimal control schemes that balance safety and efficiency.

[0006] 3) High safety risk: The monitoring and early warning of the reservoir water level are often based on a single threshold value, lacking a comprehensive evaluation of the rising trend, rising rate and future prediction value of the water level. This may lead to failure to timely initiate the most urgent flood discharge plan when the water level rapidly approaches the dangerous value. SUMMARY

[0007] In order to solve the problems of limited prediction ability, lagging response, rigid control strategy, lack of adaptability and high safety risk existing in the prior art, the present application aims to provide a reservoir flow control method and system based on artificial intelligence.

[0008] The technical solution adopted by the present application is as follows: A reservoir flow control method based on artificial intelligence, comprising the following steps: Collecting real-time multi-source monitoring data of the reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph; According to the real-time basin topology graph, using a reservoir flow and water level prediction model constructed based on a deep learning algorithm, performing prediction to obtain real-time reservoir flow and water level prediction results; If the real-time reservoir water level prediction value in the real-time reservoir flow and water level prediction results exceeds the real-time dynamic water level safety threshold value, triggering the reservoir flow control process, and proceeding to the next step; According to the real-time flow and water level prediction results, using a reservoir flow control model constructed based on a reinforcement learning algorithm, performing strategy generation to obtain a real-time reservoir flow control strategy; According to the real-time reservoir flow control strategy, generating a corresponding real-time reservoir flow control instruction for the reservoir, and executing the real-time reservoir flow control instruction based on the reservoir.

[0009] Further, the real-time multi-source monitoring data includes real-time hydrological data, real-time meteorological data, real-time remote sensing and geographic information data, and real-time reservoir operation data.

[0010] Further, collecting real-time multi-source monitoring data of the reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph, comprising the following steps: Collecting real-time multi-source monitoring data of the reservoir basin, and pre-processing the real-time multi-source monitoring data to obtain pre-processed real-time multi-source monitoring data; According to the pre-processed real-time multi-source monitoring data, discretizing the basin of the reservoir to obtain a plurality of nodes representing different spatial regions; According to the pre-processed real-time multi-source monitoring data, extracting real-time node features of each node, and constructing a real-time node feature vector of each node; According to the pre-processed real-time multi-source monitoring data, edges between nodes are determined, and a corresponding weight is set for each edge; All nodes of the reservoir basin, real-time node feature vectors, edges between nodes, and weights of each edge are coded into a graph data structure to obtain a corresponding real-time basin topology graph. The pre-processed historical multi-source monitoring data and the pre-processed real-time multi-source monitoring data of the same reservoir basin are integrated to obtain real-time time series data.

[0011] Further, the reservoir flow and water level prediction model is constructed based on the ST-ACN-PC-MLP algorithm, and the reservoir flow and water level prediction model comprises an input layer, a spatio-temporal attention layer constructed based on the CA algorithm, a spatio-temporal feature extraction layer constructed based on the ST-GCN algorithm, a context information extraction layer based on LSTM, a physical constraint layer, and a reservoir flow and water level prediction layer constructed based on the MLP algorithm.

[0012] Further, according to the real-time basin topology graph, the reservoir flow and water level prediction model constructed based on the deep learning algorithm is used for prediction to obtain real-time reservoir flow and water level prediction results, including the following steps: The real-time basin topology graph and the real-time time series data are input into the input layer of the reservoir flow and water level prediction model constructed based on the deep learning algorithm; The spatio-temporal attention layer of the reservoir flow and water level prediction model is used to generate real-time spatio-temporal attention weights, and the real-time basin topology graph and the real-time time series data are weighted and assigned according to the real-time spatio-temporal attention weights to obtain real-time weighted and assigned features; The spatio-temporal feature extraction layer of the reservoir flow and water level prediction model is used to extract real-time spatio-temporal features of the real-time weighted and assigned features; The context information extraction layer of the reservoir flow and water level prediction model is used to extract real-time context information features of the real-time spatio-temporal features; The physical constraint layer of the reservoir flow and water level prediction model is used to correct the real-time context information features to obtain physical constraint real-time context information features; The physical constraint layer of the reservoir flow and water level prediction model is used to correct the real-time context information features to obtain physical constraint real-time context information features; According to the physical constraint real-time context information features, the reservoir flow and water level prediction layer of the reservoir flow and water level prediction model is used for prediction to obtain real-time reservoir flow prediction results; the real-time reservoir flow prediction results include real-time reservoir water level prediction values and real-time reservoir flow prediction values.

[0013] Further, if the real-time reservoir flow and the real-time reservoir water level prediction value in the reservoir flow and water level prediction result exceed the real-time dynamic water level safety threshold, a reservoir flow control process is triggered, and the next step is entered, including the following steps: According to the real-time reservoir flow measured value of the real-time reservoir operation data in the pre-processed real-time multi-source monitoring data and the historical reservoir flow measured value of the historical reservoir operation data in the corresponding pre-processed historical multi-source monitoring data, a real-time flow anomaly entropy is obtained; According to the real-time reservoir water level measured value of the real-time reservoir operation data in the pre-processed real-time multi-source monitoring data and the real-time meteorological data, a real-time dynamic coefficient is obtained; According to the real-time flow anomaly entropy and the real-time dynamic coefficient, a corresponding real-time dynamic water level safety threshold is generated using a dynamic water level safety threshold calculation formula; If the real-time reservoir flow and the real-time reservoir water level prediction value in the reservoir flow and water level prediction result exceed the real-time dynamic water level safety threshold, a reservoir flow control process is triggered, and a strategy generation step is entered.

[0014] Further, the reservoir flow control model is constructed based on an ML-HRL algorithm, and the reservoir flow control model includes a high-level strategy layer and a low-level strategy layer constructed based on an HRL algorithm. The high-level strategy layer is provided with a meta-high-level strategy agent and a plurality of snapshot networks constructed based on an ML algorithm, and the low-level strategy layer includes a plurality of low-level strategy agents.

[0015] Further, according to the real-time flow and water level prediction result, a reservoir flow control model constructed based on a reinforcement learning algorithm is used for strategy generation to obtain a real-time reservoir flow control strategy, including the following steps: The real-time flow and water level prediction result and the pre-processed real-time reservoir operation data in the pre-processed real-time multi-source monitoring data are input into the reservoir flow control model constructed based on the reinforcement learning algorithm; According to the real-time flow and water level prediction result and the real-time reservoir operation data, a meta-high-level strategy agent of the high-level strategy layer is used to generate a high-level strategy, and according to the high-level strategy, a corresponding snapshot network is used to generate a plurality of low-level strategies; According to the high-level strategy and the plurality of low-level strategies, a plurality of real-time reservoir flow control decisions are generated using a plurality of low-level strategy agents of the low-level strategy layer, and all real-time reservoir flow control decisions are integrated to obtain a real-time reservoir flow control strategy.

[0016] Further, according to the real-time reservoir flow control strategy, a corresponding real-time reservoir flow control instruction of the reservoir is generated, and the real-time reservoir flow control instruction is executed based on the reservoir, including the following steps: The real-time reservoir flow control strategy is sent to a control center of the reservoir, and a plurality of real-time reservoir flow control instructions are generated according to a plurality of real-time reservoir flow control decisions in the real-time reservoir flow control strategy. The real-time reservoir flow control instructions are sent to corresponding execution equipment of the reservoir, and the real-time reservoir flow control instructions are executed by using the execution equipment.

[0017] A reservoir flow control system based on artificial intelligence is used to implement a reservoir flow control method, and the system comprises a graph structure conversion unit, a reservoir flow and water level prediction unit, a dynamic water level safety threshold determination unit, a reservoir flow control strategy generation unit, and a reservoir flow control strategy execution unit.

[0018] The beneficial effects of the present application are: The reservoir flow control method and system based on artificial intelligence provided by the present application can more comprehensively and finely depict the basin water system structure and real-time water regime by collecting real-time multi-source monitoring data (such as rainfall, water level, flow, and weather, etc.) in the basin and constructing a basin topological graph by using graph structure conversion; the reservoir flow and water level prediction model constructed based on a deep learning algorithm can learn complex hydrological laws and spatio-temporal correlations, and can more accurately predict the change trend of future reservoir flow and water level compared with traditional simplified models or empirical judgments, especially in complex and extreme conditions, the prediction accuracy and advance amount are significantly improved; the reservoir flow control strategy model constructed based on a reinforcement learning algorithm can dynamically adjust the control scheme according to the real-time prediction results, and this strategy is not a rigid fixed rule, but can autonomously learn and optimize the optimal, even phased, flood discharge or water storage strategy according to the current reservoir state, the predicted future change, and the preset safety target; when it is predicted that the water level may exceed the safety threshold, the system can automatically trigger and quickly generate an optimized control strategy, avoiding the delay and subjective bias of manual decision-making, and making the control more timely and accurate; the combination of real-time prediction and dynamic threshold judgment enables the system to more early identify potential risks, timely start the flow control process, effectively prevent the reservoir water level from being over-stored, and reduce the dam collapse risk; the whole process from data collection, prediction, decision-making to instruction execution has a high degree of automation, greatly shortens the response time, and improves the operation efficiency and intelligent level of the whole reservoir management. Other beneficial effects of the present application will be further described in the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flow block diagram of the reservoir flow control method based on artificial intelligence in the present application.

[0020] Figure 2 is a structural block diagram of the reservoir flow control system based on artificial intelligence in the present application. DETAILED DESCRIPTION

[0021] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1: like Figure 1 As shown, this embodiment provides a reservoir flow control method based on artificial intelligence, including the following steps: S1: Collect real-time multi-source monitoring data of the reservoir basin and perform graph structure transformation on the real-time multi-source monitoring data to obtain the corresponding real-time basin topology map, including the following steps: S1-1: Collect real-time multi-source monitoring data of the reservoir basin and preprocess the real-time multi-source monitoring data to obtain preprocessed real-time multi-source monitoring data; Real-time multi-source monitoring data includes real-time hydrological data, real-time meteorological data, real-time remote sensing and geographic information data, and real-time reservoir operation data; Real-time hydrological data includes watershed division, watershed flow direction, watershed flow rate, watershed flow velocity, and watershed sediment content, etc. Real-time meteorological data includes rainfall, wind speed, wind direction, temperature, humidity, and evaporation. Real-time remote sensing and geographic information data include remote sensing images covering the reservoir basin, digital elevation data, land use / cover data, and watershed river system maps, etc. Real-time reservoir operation data includes measured reservoir flow (inflow and outflow), measured reservoir water level, reservoir flow velocity, reservoir scheduling rules, flood control level, dead water level, normal storage level, operating procedures for flood discharge facilities (such as gates), and generator characteristic curves, etc. Data preprocessing includes: Data cleaning: Detecting and handling missing values ​​(e.g., using interpolation, mean imputation, etc.) and outliers (e.g., using statistical methods or filtering based on physical rules); Data alignment and interpolation: unifying data from different sources and with different time resolutions to the same time base (such as every minute or every 5 minutes), and interpolating low-frequency data (such as linear interpolation and spline interpolation). Data normalization / standardization: Normalize or standardize data of different dimensions to make their numerical range suitable for subsequent model input, such as using Min-Max normalization or Z-score standardization; Data alignment: unifying timestamps and matching spatial scales for data from different sources, with different time frequencies, and different spatial resolutions; for example, matching high-resolution meteorological data with hydrological monitoring stations through downscaling or aggregation methods, and georeferencing remote sensing data with digital elevation data. S1-2: Discretize the reservoir basin according to the basin division or grid division of digital elevation data in the pre-processed real-time multi-source monitoring data, to obtain a plurality of nodes representing different spatial regions; S1-3: Extract real-time node features of each node according to the pre-processed real-time multi-source monitoring data, and construct a real-time node feature vector of each node; Real-time node features include: 1) Real-time hydrological features: Node current measured flow: key basic data, directly reflecting the water flow state; Upstream inflow: important input, reflecting the influence of upstream inflow; Flow direction information: key indicator of spatial dynamics, which needs to be encoded as a vector or angle; Basin flow / flow direction / flow rate change rate: dynamic trend feature, enhancing prediction ability; Influence on reservoir: needs to be quantified as specific indicators (such as weighted flow of confluence time); Basin flow rate change rate: capture acceleration / deceleration trend, improve prediction accuracy; Previous impact rainfall: classic hydrological indicator, needs to be calculated in combination with soil moisture model; 2) Real-time weather features: Effective rainfall: calculated using the double mass curve method, considering the previous impact; Potential evapotranspiration: Penman-Monteith formula is recommended; Temperature and humidity: need to be converted into water vapor pressure or saturation difference, etc. The indicators with clear hydrological meaning; Wind field effect: needs to be quantified as water surface evaporation enhancement coefficient; 3) Real-time spatial features: Digital elevation features: curvature features (confluence and collection characteristics); Underlying surface runoff coefficient: dynamically updated based on soil moisture adjustment; Topological location features: sub-basin area proportion index; Topological distance: weighted distance is used to consider flow rate change; 4) Real-time time features: Event time: including monthly / decadal / flood season classification labels; Periodic features: frequency features after Fourier transform; 5) Real-time state features: Water level features: distance of current water level from flood control water level / normal storage level, proportion of current outflow to maximum discharge capacity; Flow features: inflow, outflow, reservoir flow rate; Dispatching rules: multi-label coding (such as flood control + benefit + power generation); Scheduling target: flood control weight 0.6, water supply weight 0.3, power generation weight 0.1; 6) Real-time multi-source data fusion features: Rainfall-terrain interaction (slope x rainfall intensity); Soil-vegetation interaction (humidity x vegetation coverage); Meteorological-hydrological interaction (temperature x flow rate change rate); S1-4: According to the basin flow direction, basin flow rate and basin water system map in the pre-processed real-time multi-source monitoring data, determine the edges between nodes, and set the corresponding weight for each edge; The weight of the edge can represent the difficulty, distance or topological importance of water flow transmission; S1-5: Encode all nodes of the reservoir basin, real-time node feature vectors, edges between nodes and weights of each edge into a graph data structure to obtain the corresponding real-time basin topology graph; S1-6: Integrate several pre-processed historical multi-source monitoring data and pre-processed real-time multi-source monitoring data of the same reservoir basin to obtain real-time time series data; S2: According to the real-time basin topology graph, use the reservoir flow and water level prediction model constructed based on deep learning algorithm for prediction to obtain real-time reservoir flow and water level prediction results; The reservoir flow and water level prediction model is constructed based on Spatio-Temporal Attention Convolutional Network with Physical Constraints (ST-ACN-PC) - Multi-Layer Perceptron (MLP) algorithm, and the reservoir flow and water level prediction model includes an input layer, a spatio-temporal attention layer constructed based on Coordinate Attention (CA) algorithm, a spatio-temporal feature extraction layer constructed based on Spatio-Temporal Graph Convolutional Network (ST-GCN) algorithm, a context information extraction layer based on Long Short-Term Memory (LSTM), a physical constraint layer and a reservoir flow and water level prediction layer based on MLP algorithm, which are connected in turn; According to the real-time basin topology graph, use the reservoir flow and water level prediction model constructed based on deep learning algorithm for prediction to obtain real-time reservoir flow and water level prediction results, including the following steps: S2-1: Obtain the real-time basin topology graph And real-time time series data An input layer of a reservoir flow and water level prediction model constructed based on a deep learning algorithm is input, wherein, a real-time basin topology graph is obtained, a node set, an edge set, and a weight of the edge is represented by an adjacency matrix , N a total number of historical data, t a time indication quantity; The formula is:

[0023] In the formula, an input layer feature; a node feature matrix; S2-2: using a spatio-temporal attention layer of the reservoir flow and water level prediction model, generating a real-time spatio-temporal attention weight, and according to the real-time spatio-temporal attention weight, weighting and assigning the real-time basin topology graph and the real-time time series data to obtain a real-time weighted and assigned feature; The formula is:

[0024] In the formula, a real-time weighted and assigned feature; a real-time spatio-temporal attention weight; a spatio-temporal attention generation function; a bias; S2-3: using a spatio-temporal feature extraction layer of the reservoir flow and water level prediction model, extracting a real-time spatio-temporal feature of the real-time weighted and assigned feature; The formula is:

[0025] In the formula, a real-time spatio-temporal feature; an adjacency matrix expansion; an ST-GCN weight matrix; a t-d real-time weighted and assigned feature at the moment; d a time lag; D a maximum time lag step number; S2-4: using a context information extraction layer of the reservoir flow and water level prediction model, extracting a real-time context information feature of the real-time spatio-temporal feature; The formula is:

[0026] In the formula, t t- ​​Real-time context information features and historical context information features at time 1; This is a function for extracting contextual information. S2-5: Using the physical constraint layer of the reservoir flow and water level prediction model, the real-time context information features are corrected to obtain the real-time context information features after physical constraints. The formula is:

[0027] In the formula, Features of real-time contextual information after physical constraints; This is a correction function used to check and correct states that do not conform to physical laws, adjusting characteristics according to physical equations; S2-6: Based on the real-time contextual information features after physical constraints, the reservoir flow and water level prediction layer of the reservoir flow and water level prediction model is used to make predictions and obtain real-time reservoir flow prediction results; the real-time reservoir flow prediction results include real-time reservoir water level prediction values ​​and real-time reservoir flow prediction values. The formula is:

[0028] In the formula, For the future k A sequence of real-time reservoir flow forecasts at each time step; For the future k A sequence of real-time reservoir water level predictions at each time step; For reservoir flow prediction functions and reservoir water level prediction functions; k For time step parameters; The reservoir flow and water level prediction model needs to be trained and optimized in advance. The training objective is to minimize the error between the predicted value and the actual value, while optimizing the physical constraint loss term of the physical constraint layer. This is a multi-objective optimization problem, which is usually solved by combining multiple loss terms through weighted summation. The formula for the loss function of the reservoir flow and water level prediction model is as follows:

[0029] In the formula, This is the total loss function; Hyperparameters used to control the relative importance of physical constraint losses; for t Real-time predicted reservoir flow and real-time predicted reservoir water level at any given time; The physical constraint loss function;

[0030] In the formula, The prediction error loss function; is a mean square error function; is t is a real-time reservoir flow prediction value and a real-time reservoir water level prediction value at a time point; is a flow prediction error weight coefficient and a water level prediction error weight coefficient; is t is a real-time reservoir flow measured value and a real-time reservoir water level measured value at a time point; T is a total number of times;

[0031] In the formula, is a physical constraint loss function; is a first and second physical constraint loss weight coefficient; is a water level threshold constraint function, representing that the water level cannot exceed a certain threshold; is a water balance constraint function; is t+i is a real-time reservoir water level prediction value and a reservoir water level threshold at a time point; i is a time step indication; is a water level and storage capacity curve calculation value; is a real-time reservoir inflow and a real-time reservoir outflow in a real-time reservoir flow measured value; is an evaporation amount; S3: if the real-time reservoir flow and the real-time reservoir water level prediction value in the real-time reservoir water level prediction result exceed the real-time dynamic water level safety threshold, triggering a reservoir flow control process, and entering the next step, including the following steps: S3-1: according to the real-time reservoir flow measured value in the real-time reservoir operation data of the pre-processed real-time multi-source monitoring data and the historical reservoir flow measured value in the historical reservoir operation data of the corresponding several pre-processed historical multi-source monitoring data, obtaining a real-time flow anomaly entropy; The formula is:

[0032] In the formula, is a historical k absolute difference of flow in a time step; k is a time step parameter; is t is a real-time reservoir flow measured value at a time point; is a historical k reservoir flow measured value in a time step;

[0033] In the formula, is a real-time flow anomaly entropy; is a probability distribution of absolute difference of flow; is the total number of time steps; S3-2: According to the real-time reservoir water level measured value of the real-time reservoir operation data in the pre-processed real-time multi-source monitoring data and the real-time meteorological data, a real-time dynamic coefficient is obtained; The formula is:

[0034] In the formula, is the real-time dynamic coefficient; is the first and second calculation weight coefficient; is t is the real-time reservoir water level measured value at the moment; is the real-time meteorological data; S3-3: According to the real-time flow anomaly entropy and the real-time dynamic coefficient, a corresponding real-time dynamic water level safety threshold is generated using a dynamic water level safety threshold calculation formula; The formula is:

[0035] In the formula, is the real-time dynamic water level safety threshold; is the first and second weight parameter; is the basic water level safety threshold; S3-4: If the real-time reservoir flow and the real-time reservoir water level predicted value in the real-time flow and water level prediction result exceed the real-time dynamic water level safety threshold, the reservoir flow control process is triggered, and the strategy generation step is entered; S4: According to the real-time flow and water level prediction result, a reservoir flow control model constructed based on a reinforcement learning algorithm is used to generate a strategy, and a real-time reservoir flow control strategy is obtained; The reservoir flow control model is constructed based on a Meta-Learning (ML)-Hierarchical Reinforcement Learning (HRL) algorithm, and the reservoir flow control model includes a high-level strategy layer and a low-level strategy layer constructed based on the HRL algorithm. The high-level strategy layer is provided with a meta-high-level strategy agent and a plurality of snapshot networks constructed based on the ML algorithm, and the low-level strategy layer includes a plurality of low-level strategy agents; According to the real-time flow and water level prediction result, a reservoir flow control model constructed based on a reinforcement learning algorithm is used to generate a strategy, and a real-time reservoir flow control strategy is obtained, including the following steps: S4-1: The real-time flow and water level prediction result and the pre-processed real-time reservoir operation data in the pre-processed real-time multi-source monitoring data are input into the reservoir flow control model constructed based on the reinforcement learning algorithm; S4-2: According to the real-time flow and water level prediction results and real-time reservoir operation data, using the meta-high-level strategy agent of the high-level strategy layer, a high-level strategy is generated, and according to the high-level strategy, using the corresponding snapshot network, a plurality of low-level strategies are generated; The formula is:

[0036] In the formula, is a high-level strategy, for example, an abstract action plan or goal (such as "lower water level", "maintain current outflow", "flood control mode", "power generation mode", "ecological mode", etc.); is a strategy function of the meta-high-level strategy agent; is a state vector constructed from real-time flow and water level prediction results and real-time reservoir operation data;

[0037] In the formula, is the low-level strategy output by the snapshot network; is the low-level strategy output by the snapshot network, which is more specific than the high-level strategy, for example, "floodgate A is opened at a certain rate for 1 hour in the future", "floodgate B is opened at another rate for 2 hours in the future; is a strategy function of the snapshot network; is a snapshot network indicator; S4-3: According to the high-level strategy and a plurality of low-level strategies, using a plurality of low-level strategy agents of the low-level strategy layer, a plurality of real-time reservoir flow control decisions are generated, and all real-time reservoir flow control decisions are integrated to obtain a real-time reservoir flow control strategy; The formula is:

[0038] In the formula, is the real-time reservoir flow control decision output by the low-level strategy agent; is a specific control instruction, for example, "floodgate A is opened by 5% in the next control period", "floodgate B is opened by 3% in the next control period"; is a strategy function of the low-level strategy agent; is a low-level strategy agent indicator; is the low-level strategy received by the low-level strategy agent; is the low-level strategy received by the low-level strategy agent; S5: According to the real-time reservoir flow control strategy, a real-time reservoir flow control instruction corresponding to the reservoir is generated, and the real-time reservoir flow control instruction is executed based on the reservoir, including the following steps: S5-1: Send the real-time reservoir flow control strategy to the reservoir control center, and generate several real-time reservoir flow control instructions (such as specifying the gate number, target opening degree, execution time, etc.) based on several real-time reservoir flow control decisions in the real-time reservoir flow control strategy. S5-2: Send the real-time reservoir flow control command to the corresponding execution device of the reservoir, and use the execution device to execute the real-time reservoir flow control command; after receiving the command, the execution device drives the gate, valve and other physical components to perform actions according to the command requirements (such as opening, closing, adjusting the opening degree), change the flood discharge flow of the reservoir, and thus control the reservoir water level within a safe range.

[0039] Example 2: like Figure 2 As shown, this embodiment provides an artificial intelligence-based reservoir flow control system for implementing a reservoir flow control method. The system includes a graph structure conversion unit, a reservoir flow and water level prediction unit, a dynamic water level safety threshold determination unit, a reservoir flow control strategy generation unit, and a reservoir flow control strategy execution unit.

[0040] The graph structure conversion unit is used to collect real-time multi-source monitoring data of the reservoir basin and perform graph structure conversion on the real-time multi-source monitoring data to obtain the corresponding real-time basin topology map. The reservoir flow and water level prediction unit is used to predict the reservoir flow and water level based on a deep learning algorithm-based prediction model according to the real-time watershed topology map, and obtain the real-time reservoir flow and water level prediction results. The dynamic water level safety threshold determination unit is used to trigger the reservoir flow control process and proceed to the next step if the real-time reservoir flow and the real-time reservoir water level prediction value in the water level prediction result exceed the real-time dynamic water level safety threshold. The reservoir flow control strategy generation unit is used to generate a real-time reservoir flow control strategy based on a reservoir flow control model constructed using a reinforcement learning algorithm, based on real-time flow and water level prediction results. The reservoir flow control strategy execution unit is used to generate real-time reservoir flow control instructions corresponding to the reservoir based on the real-time reservoir flow control strategy, and to execute the real-time reservoir flow control instructions based on the reservoir.

[0041] The application provides a reservoir flow control method and system based on artificial intelligence, which can more comprehensively and finely depict the river basin water system structure and real-time water regime by collecting real-time multi-source monitoring data (such as rainfall, water level, flow, weather, etc.) in the river basin and constructing a river basin topology graph by using a graph structure conversion; the reservoir flow and water level prediction model constructed based on a deep learning algorithm can learn complex hydrological laws and spatio-temporal correlations, and can more accurately predict the future change trend of reservoir flow and water level compared with traditional simplified models or experience judgments, especially in complex and extreme conditions, the prediction accuracy and advance quantity are significantly improved; the reservoir flow control strategy model constructed based on a reinforcement learning algorithm can dynamically adjust the control scheme according to the real-time prediction result, and the strategy is not a rigid fixed rule, but can autonomously learn and optimize the optimal, even phased, flood discharge or water storage strategy according to the current reservoir state, the predicted future change and the preset safety target; when it is predicted that the water level may exceed the safety threshold, the system can automatically trigger and quickly generate an optimized control strategy, avoiding the delay and subjective bias of manual decision-making, so that the control is more timely and accurate; the combination of real-time prediction and dynamic threshold judgment enables the system to more early identify potential risks and timely start the flow control process, effectively preventing reservoir water level over storage and reducing the risk of dam collapse; the whole process from data collection, prediction, decision-making to instruction execution has high automation degree, greatly shortens the response time and improves the operation efficiency and intelligent level of the whole reservoir management.

[0042] The application is not limited to the above optional embodiments, and anyone can derive other various forms of products under the inspiration of the application. The above specific embodiments should not be understood as limiting the protection scope of the application, and the protection scope of the application should be defined by the claims, and the specification can be used to explain the claims.

Claims

1. An artificial intelligence-based reservoir flow control method, characterized by: The method comprises the following steps: Collecting real-time multi-source monitoring data of a reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph; According to the real-time basin topology graph, using a reservoir flow and water level prediction model constructed based on a deep learning algorithm to perform prediction to obtain real-time reservoir flow and water level prediction results; If the real-time reservoir water level prediction value in the real-time reservoir flow and water level prediction results exceeds the real-time dynamic water level safety threshold, triggering a reservoir flow control process, and entering the next step; According to the real-time flow and water level prediction results, using a reservoir flow control model constructed based on a reinforcement learning algorithm to perform strategy generation to obtain a real-time reservoir flow control strategy; According to the real-time reservoir flow control strategy, generating a corresponding real-time reservoir flow control instruction of the reservoir, and executing the real-time reservoir flow control instruction based on the reservoir.

2. The reservoir flow control method based on artificial intelligence according to claim 1, characterized in that: The real-time multi-source monitoring data includes real-time hydrological data, real-time meteorological data, real-time remote sensing and geographic information data, and real-time reservoir operation data.

3. The reservoir flow control method based on artificial intelligence according to claim 2, characterized in that: Collecting real-time multi-source monitoring data of a reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph, comprising the following steps: Collecting real-time multi-source monitoring data of a reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph, comprising the following steps: Collecting real-time multi-source monitoring data of a reservoir basin, and performing graph structure conversion on the real-time multi-source monitoring data to obtain a corresponding real-time basin topology graph, comprising the following steps: According to the preprocessed real-time multi-source monitoring data, extracting real-time node features of each node, and constructing a real-time node feature vector of each node; According to the preprocessed real-time multi-source monitoring data, determining the edges between the nodes, and setting the corresponding weights for each edge; Encoding all nodes of the reservoir basin, real-time node feature vectors, edges between nodes, and weights of each edge into a graph data structure to obtain a corresponding real-time basin topology graph; Integrating a plurality of preprocessed historical multi-source monitoring data and preprocessed real-time multi-source monitoring data of the same reservoir basin to obtain real-time time series data.

4. The reservoir flow control method based on artificial intelligence according to claim 3, characterized in that: The reservoir flow and water level prediction model is constructed based on an ST-ACN-PC-MLP algorithm, and the reservoir flow and water level prediction model comprises an input layer, a spatio-temporal attention layer constructed based on a CA algorithm, a spatio-temporal feature extraction layer constructed based on an ST-GCN algorithm, an LSTM context information extraction layer, a physical constraint layer, and a reservoir flow and water level prediction layer constructed based on an MLP algorithm.

5. The reservoir flow control method based on artificial intelligence according to claim 4, characterized in that: According to the real-time basin topology graph, using a reservoir flow and water level prediction model constructed based on a deep learning algorithm to perform prediction to obtain real-time reservoir flow and water level prediction results, comprising the following steps: Inputting the real-time basin topology graph and the real-time time series data into the input layer of the reservoir flow and water level prediction model constructed based on the deep learning algorithm; Using the spatio-temporal attention layer of the reservoir flow and water level prediction model to generate real-time spatio-temporal attention weights, and weighting and assigning the real-time basin topology graph and the real-time time series data according to the real-time spatio-temporal attention weights to obtain real-time weighted and assigned features; The real-time spatio-temporal feature of the real-time weighted assignment feature is extracted by using the spatio-temporal feature extraction layer of the reservoir flow and water level prediction model; The real-time context information feature of the real-time spatio-temporal feature is extracted by using the context information extraction layer of the reservoir flow and water level prediction model; The real-time reservoir flow prediction result is obtained by using the reservoir flow and water level prediction layer of the reservoir flow and water level prediction model according to the real-time context information feature after the physical constraint; 6. The reservoir flow control method based on artificial intelligence according to claim 5, characterized in that: If the real-time reservoir water level prediction value in the real-time reservoir flow and water level prediction result exceeds the real-time dynamic water level safety threshold, the reservoir flow control process is triggered, and the next step is entered, including the following steps: The real-time flow anomaly entropy is obtained according to the real-time reservoir flow measured value of the real-time reservoir operation data in the preprocessed real-time multi-source monitoring data and the historical reservoir flow measured value of the historical reservoir operation data in the corresponding preprocessed historical multi-source monitoring data; The real-time dynamic coefficient is obtained according to the real-time reservoir water level measured value of the real-time reservoir operation data in the preprocessed real-time multi-source monitoring data and the real-time meteorological data; The corresponding real-time dynamic water level safety threshold is generated by using the dynamic water level safety threshold calculation formula according to the real-time flow anomaly entropy and the real-time dynamic coefficient; If the real-time reservoir water level prediction value in the real-time reservoir flow and water level prediction result exceeds the real-time dynamic water level safety threshold, the reservoir flow control process is triggered, and the strategy generation step is entered.

7. The reservoir flow control method based on artificial intelligence according to claim 6, characterized in that: The reservoir flow control model is constructed based on the ML-HRL algorithm, and the reservoir flow control model includes a high-level strategy layer and a low-level strategy layer constructed based on the HRL algorithm, the high-level strategy layer is provided with a meta-high-level strategy agent and a plurality of snapshot networks constructed based on the ML algorithm, and the low-level strategy layer includes a plurality of low-level strategy agents.

8. The reservoir flow control method based on artificial intelligence according to claim 7, characterized in that: According to the real-time flow and water level prediction result, the reservoir flow control model constructed based on the reinforcement learning algorithm is used to generate a strategy, and a real-time reservoir flow control strategy is obtained, including the following steps: The real-time flow and water level prediction result and the preprocessed real-time reservoir operation data in the preprocessed real-time multi-source monitoring data are input into the reservoir flow control model constructed based on the reinforcement learning algorithm; According to the real-time flow and water level prediction result and the real-time reservoir operation data, the meta-high-level strategy agent of the high-level strategy layer is used to generate a high-level strategy, and according to the high-level strategy, the corresponding snapshot network is used to generate a plurality of low-level strategies; According to the high-level strategy and the plurality of low-level strategies, the plurality of low-level strategy agents of the low-level strategy layer are used to generate a plurality of real-time reservoir flow control decisions, and all real-time reservoir flow control decisions are integrated to obtain a real-time reservoir flow control strategy.

9. The reservoir flow control method based on artificial intelligence according to claim 8, characterized in that: According to the real-time reservoir flow control strategy, the corresponding real-time reservoir flow control instruction of the reservoir is generated, and the real-time reservoir flow control instruction is executed based on the reservoir, including the following steps: The real-time reservoir flow control strategy is sent to the control center of the reservoir, and a plurality of real-time reservoir flow control instructions are generated according to the plurality of real-time reservoir flow control decisions in the real-time reservoir flow control strategy; The real-time reservoir flow control instruction is sent to a corresponding execution device of the reservoir, and the real-time reservoir flow control instruction is executed by using the execution device.

10. An artificial intelligence-based reservoir flow control system for implementing the reservoir flow control method according to any one of claims 1-9, characterized by: The system comprises a graph structure conversion unit, a reservoir flow and water level prediction unit, a dynamic water level safety threshold determination unit, a reservoir flow control strategy generation unit and a reservoir flow control strategy execution unit.

Images