An AI hydrological prediction-based watershed cascade reservoir joint scheduling optimization system

Abstract

The application discloses a kind of based on AI hydrology prediction watershed cascade reservoir joint scheduling optimization system, belong to watershed cascade reservoir joint scheduling technical field.The system is connected by data fusion, hydrology prediction, collaborative scheduling and scheduling correction module in turn communication connection composition.Data fusion module processes multiple-source hydrology meteorological data to generate standardization dataset;Hydrology prediction module constructs the AI dynamic prediction model of fusion of LSTM and GRU, combines dynamic error correction and extreme scene reinforcement learning, outputs future 96 hours time-to-time hydrology prediction result and confidence;Collaborative scheduling module establishes dynamic scheduling rule base and upstream and downstream water power time difference grading compensation mechanism, and distributes scheduling authority according to flood control> ecology> power generation priority;Scheduling correction module generates regular and extreme scene differentiation instruction, and is based on execution deviation grading dynamic correction.The application solves the problems of low precision of traditional hydrology prediction, cascade reservoir collaborative response lag, realizes multi-objective overall optimal scheduling, and significantly improves water resource utilization efficiency and watershed flood control safety.

Description

Technical Field

[0001] This invention belongs to the field of joint scheduling technology of cascade reservoirs in a river basin, and specifically relates to a joint scheduling optimization system for cascade reservoirs in a river basin based on AI hydrological prediction. Background Technology

[0002] Under the influence of global climate change, watershed hydrological processes exhibit nonlinearity, abrupt changes, and uncertainty, making traditional hydrological forecasting methods insufficient to meet the accuracy and timeliness requirements of modern water resource management. Particularly in cascade reservoir systems, the interconnectedness of water volumes in upstream and downstream reservoirs and the mutual influence of their scheduling behaviors lead to amplified inflow forecasting errors at each level, resulting in decreased reliability of scheduling plans and impacting the overall balance between flood control, water supply security, and power generation benefits. Current cascade reservoir scheduling generally suffers from insufficient forecasting accuracy, weak inter-reservoir coordination, and limited real-time decision-making capabilities, often relying on experience-based judgments and failing to achieve overall basin-wide optimization. With the development of artificial intelligence technology, hydrological forecasting methods based on deep learning and multi-source data fusion can effectively improve the forecasting capabilities of runoff and flood processes, providing a new technical approach for the intelligent joint scheduling of cascade reservoirs.

[0003] A patent with publication number CN119398419B discloses a cross-basin multi-station hydropower station centralized control and optimization platform, comprising: a data acquisition unit that uses a variable acquisition frequency mechanism to collect hydrological data, power grid load data, and environmental meteorological data; a hydrological monitoring unit that analyzes and monitors cross-basin hydrological dynamics in real time based on hydrological data and uses a hydrological prediction model to predict cross-basin hydrological changes; a hydrological scheduling simulation unit that simulates hydrological scheduling based on the predicted cross-basin hydrological changes using the hydrological prediction model, and constructs a scheduling strategy; a water resources scheduling unit that constructs an adaptive global optimization scheduling framework and generates the optimal scheduling strategy; and an emergency scheduling unit that handles water resources scheduling in emergency situations. This cross-basin multi-station hydropower station centralized control and optimization platform, through an adaptive global optimization scheduling framework combined with a hydrological prediction model, dynamically adjusts the hydropower station's scheduling strategy in real time and centrally controls and optimizes the stable operation of cross-basin water resources.

[0004] Although an existing cross-basin multi-station hydropower station centralized control and optimization platform collects hydrological, power grid load and environmental meteorological data through IoT sensor networks, combines a hydrological prediction model constructed with a long short-term memory network to predict cross-basin hydrological changes, and then simulates and adjusts the hydropower station scheduling strategy through an adaptive global optimization scheduling framework to achieve preliminary allocation and scheduling optimization of cross-basin water resources, the low prediction accuracy and lack of dynamic collaborative scheduling mechanism among cascade reservoirs in the basin lead to slow joint scheduling response and low collaborative efficiency. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction. By constructing an adaptive, high-precision AI dynamic hydrological prediction model and integrating multi-source data, it effectively solves the problems of insufficient capture of complex hydrological features and high prediction uncertainty in traditional methods. At the same time, it establishes a collaborative and dynamically correctable scheduling decision system, overcoming problems such as inaccurate time difference compensation, conflicting joint scheduling instructions, and delayed emergency response in cascade reservoir linkage, thereby comprehensively improving the water resource utilization efficiency and flood control safety guarantee capabilities of basin-wide cascade reservoirs.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A watershed cascade reservoir joint scheduling optimization system based on AI hydrological prediction, comprising the following steps: It includes a data fusion module, a hydrological prediction module, a collaborative scheduling module, and a scheduling correction module, which are connected in sequence: The data fusion module is used to acquire real-time monitoring data, real-time meteorological data and historical hydrological data in the watershed, and after preprocessing, generate a multi-source fusion dataset and output it to the hydrological prediction module. The hydrological prediction module is used to receive the multi-source fusion dataset, construct an AI dynamic hydrological prediction model, automatically identify the weights of key hydrological factors under different climatic conditions, iteratively optimize model parameters using an incremental learning mechanism, introduce a confidence assessment mechanism to construct a hydrological error feedback function to dynamically correct prediction bias, trigger reinforcement learning branch training for extreme hydrological scenarios, and output hydrological prediction results and hydrological confidence to the collaborative scheduling module. The collaborative scheduling module is used to receive the hydrological prediction results and hydrological confidence levels, construct and dynamically update the joint hydrological scheduling rule base, optimize the core scheduling parameters by combining hydrological confidence levels and hydrological change trends, and configure the control mode according to the confidence level; generate a watershed hydraulic topology map through topology analysis, quantify the response time difference and impact of upstream and downstream scheduling, establish time difference graded compensation rules to compensate for hydraulic transmission delays; introduce a distributed collaborative decision-making mechanism to realize real-time data sharing and dynamic allocation of scheduling permissions, and output the optimal scheduling rule set, time difference compensation scheme, and scheduling permission allocation results to the scheduling correction module; The scheduling correction module is used to receive the optimal scheduling rule set, time difference compensation scheme and scheduling authority allocation results, generate a first scheduling instruction adapted to conventional hydrological scenarios and a second scheduling instruction adapted to extreme hydrological scenarios, and integrate them to form a basin-level cascade reservoir scheduling instruction set; obtain scheduling execution feedback data in real time to calculate comprehensive scheduling deviation, determine the deviation level and trigger graded response, and dynamically correct differentiated scheduling instructions.

[0007] Preferably, the preprocessing process of the data fusion module is as follows: A wavelet threshold denoising algorithm is used to filter electromagnetic interference noise in real-time monitoring data, abnormal extreme value data is removed based on the 3σ criterion, and time sequence alignment of multi-site data is achieved through a timestamp synchronization algorithm. A feature alignment and fusion algorithm is used to associate and preprocessed real-time monitoring data, real-time meteorological data, and historical hydrological data to generate a standardized multi-source fusion dataset. The multi-source fusion dataset includes instantaneous rainfall intensity, real-time reservoir water level, inflow / outflow, meteorological forecast data, historical runoff data, and static parameters of cascade reservoirs.

[0008] Preferably, the specific steps for the hydrological prediction module to construct the AI ​​dynamic hydrological prediction model are as follows: Based on the multi-source fusion dataset, an AI dynamic hydrological prediction model is constructed, which includes an input layer, a feature extraction layer, a parallel fusion feature processing layer, and an output layer. The input layer receives a multi-source fusion dataset, matches the corresponding key hydrological factors according to the climate type, extracts time-series data, and outputs a multi-dimensional feature vector. The feature extraction layer receives the multidimensional feature vectors and groups them according to climate type. It uses an improved Relief-F algorithm to calculate the difference between the dissimilarity of each feature in the nearest neighbor samples and the dissimilarity of the same type of nearest neighbor. It then calculates the associated feature weights by combining the climate weight factor with the initial weights. If the associated feature weight is greater than a preset screening threshold, it is determined to be a key hydrological feature and is weighted to output a weighted key feature matrix. Otherwise, it is determined to be a redundant feature and is temporarily stored in the feature pool. The parallel fusion feature processing layer constructs parallel processing units according to the first preset ratio. The first processing unit extracts long-period feature vectors from time slices with a step size of 24 using LSTM, and the second processing unit extracts short-period feature vectors from time slices with a step size of 1 using GRU. After performing dual-path feature extraction on the weighted key feature matrix, the weighted sum is obtained to obtain the hydrological fusion feature vector. The output layer receives the hydrological fusion feature vector, maps it to the initial hydrological prediction result through a fully connected network, and outputs the preliminary confidence correlation value. Historical hydrological measurement datasets are obtained, and the multi-source fusion dataset is divided into training and validation sets according to a second preset ratio. A joint loss function and Adam optimizer are used to train an AI dynamic hydrological prediction model.

[0009] Preferably, the specific steps for the hydrological prediction module to output the hydrological prediction results and hydrological confidence levels are as follows: By calling up historical hydrological data from the same period and comparing them with the initial hydrological prediction results, a hydrological error feedback function based on mean square error and Nash efficiency coefficient is constructed, and a dynamic correction factor is set to obtain the corrected hydrological prediction results. The hydrological confidence level is calculated by subtracting the comprehensive hydrological error from 10, and the high-confidence prediction results are determined by combining the pre-set confidence threshold. Set extreme scenario trigger thresholds θ 1. θ 2. θ 3. If the inflow in the revised hydrological forecast is greater than... θ 1 or the rate of water level change is greater than θ 2 or peak runoff greater than θ 3. Then the reinforcement learning branch is triggered, which retrieves multi-source data of similar events and the optimal hydrological prediction parameters as special training samples for training, and outputs the hydrological prediction results for extreme scenarios. By integrating the corrected hydrological prediction results with those for extreme scenarios, the error in the calculation of measured hydrological values ​​is obtained. For time periods exceeding the preset error threshold, a hydrological error feedback function is used for re-correction. Finally, the hydrological prediction results and hydrological confidence levels for each time period in the next 96 hours are output.

[0010] Preferably, the collaborative scheduling module includes a rule optimization unit, a time difference compensation unit, and a collaborative decision-making unit: The rule optimization unit is used to receive the hydrological prediction results and hydrological confidence levels, construct and dynamically update the hydrological joint scheduling rule base in combination with hydrological change trends, quantify the core scheduling parameters, configure differentiated control modes for prediction results with different confidence levels, and generate the optimal scheduling rule set. The time difference compensation unit is used to generate a watershed hydraulic topology map and mark the upstream and downstream relationships of cascade reservoirs, retrieve historical scheduling data to quantify the response time difference and impact of upstream and downstream scheduling, establish multi-level time difference hierarchical compensation rules to compensate for hydraulic transmission delay, and output a time difference compensation scheme. The collaborative decision-making unit is used to establish a real-time data sharing channel for encrypted cascade reservoirs, introduce a distributed collaborative decision-making mechanism to update the operating status and forecast data of each reservoir node, dynamically allocate scheduling permissions and execution priorities based on the demand priority of flood control > ecology > power generation, and output the scheduling permission allocation results.

[0011] Preferably, the specific steps for the rule optimization unit to construct the hydrological joint scheduling rule base are as follows: For the three types of indicators in the hydrological forecast results—inflow, water level change, and peak runoff—the difference between the current period and the previous period is calculated, and the trend of each indicator is determined based on the corresponding threshold. The three types of indicators are assigned corresponding weights. Based on the product of the trend voting results of each indicator and its corresponding weight, the overall hydrological trend is determined to be rising, stable, or falling. Based on the hydrological change trend, a trend coefficient is set, and the corrected water storage level threshold, ecological benchmark guarantee value and flood discharge flow adjustment weight are calculated in combination with the initial scheduling parameters. A rigid threshold control is applied to high-confidence prediction results, while a flexible space for first-level and second-level adjustments is configured for low-confidence prediction results, thereby generating an optimal scheduling rule set. The optimal scheduling rule set is stored according to trend type, confidence level and parameter range, and historical rules are integrated to build a hydrological joint scheduling rule base.

[0012] Preferably, the specific steps for the time difference compensation unit to quantify the upstream and downstream scheduling response time difference and its impact magnitude are as follows: The geographical coordinates of each reservoir and the connection relationship between upstream and downstream river channels are transformed into a directed graph structure, with reservoir nodes as vertices and upstream and downstream river channel connections as directed edges, to generate a watershed hydraulic topology graph. Extract the first time point of a single flood discharge / water storage action of the upstream reservoir, match it with the second time point where the water level / flow rate change of the downstream reservoir exceeds a preset threshold, and calculate the difference to obtain the hydraulic response time difference under a single scenario. C ; Historical data were grouped according to hydrological trend type, and the correlation coefficient between the upstream scheduling action adjustment and the downstream water level / flow change was calculated for each group of scenarios. The distribution of hydraulic response time difference in each scenario was statistically analyzed. The 25th and 75th percentiles were selected to divide the time difference into three intervals. The mean downstream change corresponding to the unit scheduling action adjustment in each scenario was calculated as the quantitative value of the impact. Establish time zone compensation rules: If C If the time is ≤1 hour, Level 1 compensation will be applied; if the time is <1 hour C If the time is ≤3 hours, a secondary compensation will be applied. C If the time exceeds 3 hours, a three-level compensation will be applied.

[0013] Preferably, the specific steps for the scheduling correction module to generate the scheduling instruction set are as follows: For conventional hydrological scenarios, the weights of flood control, water storage, and power generation are dynamically allocated based on the optimal scheduling rule base. The target water level for each reservoir during a given time period is calculated by combining the current water level, the predicted inflow, and the reservoir capacity. The baseline value of the outflow is determined based on the basic outflow ratio coefficient, the response time difference compensation coefficient, and the adjustment coefficient. The target power generation value is calculated by interpolation based on the target water level and the power output curve, thus forming the first scheduling instruction. In response to extreme hydrological scenarios, emergency dispatch rules are triggered and flood control priority is increased. Emergency parameters from similar historical cases are extracted to determine the upper limit of emergency water level and flow rate and the range of power generation output reduction. A dynamic emergency adjustment interface is set up to generate a second dispatch instruction. The first and second scheduling instructions are integrated according to timestamps, the coordination of cross-reservoir scheduling instructions is verified, and conflicting instructions are corrected based on the time difference compensation scheme to generate a set of scheduling instructions for cascade reservoirs in the basin.

[0014] Preferably, the specific steps for the scheduling correction module to determine the deviation level and perform dynamic correction are as follows: Real-time acquisition of execution feedback data of scheduling instruction set, and calculation of the absolute deviation between the actual values ​​of three types of indicators—water level, flow rate, and output—and the benchmark target values ​​for each time period and each reservoir; Assign corresponding weights to the three types of deviations, and calculate the ratio of the weighted sum of deviations to the baseline weighted sum to obtain the comprehensive scheduling deviation ΔG; Set deviation thresholds γ1 and γ2. If ΔG≤γ1, it is judged as a first-level deviation; if γ1<ΔG≤γ2, it is judged as a second-level deviation; if ΔG>γ2, it is judged as a third-level deviation. Based on the deviation level and scenario type, a graded response is triggered: In a normal hydrological scenario, a first-level deviation only records data, a second-level deviation initiates a normal correction process, and a third-level deviation temporarily activates the second dispatch command emergency interface; In an extreme hydrological scenario, a first-level deviation shortens the feedback cycle, a second-level deviation fine-tunes emergency parameters, and a third-level deviation immediately triggers a shutdown and release command. The scheduling parameters are dynamically adjusted based on the correction strategy, and the corrected instructions are synchronized to all cascade reservoir nodes.

[0015] A watershed-level cascade reservoir joint scheduling optimization method based on AI hydrological prediction includes the following steps: S1. Data Fusion: Acquire multi-source hydrological and meteorological data within the watershed, and generate a multi-source fused dataset after preprocessing; S2, AI hydrological prediction: Based on the multi-source fusion dataset, an AI dynamic hydrological prediction model is constructed. Through adaptive feature extraction, dual-path time series modeling and dynamic error correction, the hydrological prediction results and hydrological confidence are output. S3. Collaborative Scheduling: Based on the hydrological prediction results and hydrological confidence level, a dynamic hydrological joint scheduling rule base is constructed, the upstream and downstream hydraulic response time difference is quantified and a hierarchical compensation mechanism is established, and scheduling permissions are allocated through distributed collaborative decision-making. S4. Scheduling Correction and Execution: Generate differentiated scheduling instruction sets for normal and extreme scenarios, monitor execution feedback in real time and calculate comprehensive scheduling deviations, trigger hierarchical response to dynamically correct scheduling instructions, and realize joint optimization scheduling of cascade reservoirs in the basin.

[0016] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of a watershed cascade reservoir joint scheduling optimization method based on AI hydrological prediction.

[0017] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the watershed cascade reservoir joint scheduling optimization method based on AI hydrological prediction as described above.

[0018] The present invention can achieve the following beneficial effects: 1. Addressing the core issues of low prediction accuracy and lack of dynamic collaborative scheduling mechanisms for cascade reservoirs in existing technologies, this invention utilizes multi-source data fusion and the construction of an AI dynamic hydrological prediction model to accurately identify key hydrological factors under different climatic conditions. It dynamically corrects the prediction bias of the AI ​​dynamic hydrological prediction model by combining incremental learning and reinforcement learning mechanisms, effectively improving hydrological prediction accuracy and confidence. Based on the high-precision prediction results and hydrological change trends, it dynamically optimizes core scheduling parameters, configures differentiated control modes for high-confidence and low-confidence prediction results, and establishes time difference compensation rules to balance the hydraulic transmission delay between upstream and downstream areas, forming a full-process dynamic collaborative scheduling system. This effectively solves the response lag problem in traditional scheduling and significantly improves the real-time performance and adaptability of joint scheduling of cascade reservoirs.

[0019] 2. To address the problem of low coordination efficiency in traditional cascade reservoirs in river basins, this invention establishes a full-process scheduling deviation monitoring, hierarchical judgment, and dynamic correction mechanism. It captures deviations in three core indicators—water level, flow rate, and output—in real time and triggers differentiated response strategies based on deviation levels and scenario types. By building an encrypted real-time data sharing channel and a distributed collaborative decision-making mechanism, it achieves real-time synchronization of the operational status of each reservoir node with the prediction data from the AI ​​dynamic hydrological prediction model. Combined with dynamic allocation of scheduling permissions based on multi-objective priorities, it significantly improves the efficiency of joint scheduling. Simultaneously, it adopts differentiated control measures for fine-tuning in conventional hydrological scenarios and emergency compensation in extreme hydrological scenarios, ensuring that scheduling instructions are adaptable to various complex operating conditions. While improving coordination efficiency, it maximizes the protection of the operation of cascade reservoirs in the river basin and downstream flood control and ecological safety. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of a basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction. Figure 2 This is a flowchart of the AI ​​dynamic hydrological prediction model constructed in this invention; Figure 3 This is a flowchart of the process for outputting hydrological prediction results and hydrological confidence levels in this invention; Figure 4 This is a flowchart illustrating the process of compensating for hydraulic transmission delays between upstream and downstream reservoirs in this invention. Figure 5 This is a flowchart for determining the deviation level in this invention. Detailed Implementation

[0021] Example 1: refer to Figures 1 to 5 As shown in the figure, this embodiment introduces a basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction, including: a data fusion module, a hydrological prediction module, a collaborative scheduling module, and a scheduling correction module; The data fusion module is used to acquire real-time monitoring data from rain gauges, water level stations, and hydrological stations within the basin via IoT terminals. It uses encrypted transmission protocols to acquire real-time meteorological data and distributed data crawling technology to acquire historical hydrological data in batches. Preprocessing of the real-time monitoring data is performed at the edge, including: filtering electromagnetic interference noise using wavelet threshold denoising algorithms, removing outlier data based on the 3σ criterion, and synchronizing the time sequence of data from multiple stations using timestamp synchronization algorithms. In the cloud, a feature alignment fusion algorithm is used to associate and label the preprocessed multi-source data, generating a multi-source fusion dataset. The real-time monitoring data includes instantaneous rainfall intensity, real-time reservoir water level, and inflow / outflow. The multi-source data includes real-time monitoring data, real-time meteorological data, historical hydrological data, and static data from cascade reservoirs. The static data from cascade reservoirs includes reservoir capacity and flood control limit water level.

[0022] Example 2: This embodiment provides a detailed explanation of the hydrological prediction module based on Embodiment 1.

[0023] The hydrological prediction module is used to build an AI dynamic hydrological prediction model based on a multi-source fusion dataset, combining a deep long short-term memory network and a gated recurrent unit fusion architecture. An adaptive feature extraction module is embedded to automatically identify the weights of key hydrological factors under different climatic conditions. An incremental learning mechanism is used to iteratively optimize the AI ​​dynamic hydrological prediction model parameters by incorporating new data every hour, enhancing its ability to track and adapt to dynamic water inflow processes. A prediction result confidence assessment mechanism is introduced, which compares the actual hydrological data of the same period in history with the prediction results of the AI ​​dynamic hydrological prediction model, constructing a hydrological error feedback function based on mean square error and hydrological error to dynamically correct prediction biases. For extreme hydrological scenarios, a reinforcement learning branch of the AI ​​dynamic hydrological prediction model is triggered, calling similar event data from historical case sets for specialized training to ensure the prediction stability of the AI ​​dynamic hydrological prediction model in complex scenarios. Finally, it outputs high-precision hydrological prediction results and hydrological confidence levels for each time period of the next 96 hours. The hydrological prediction results include inflow, water level changes, and peak runoff. Preferably, the specific steps for constructing an AI dynamic hydrological prediction model include: Based on a multi-source fusion dataset, an AI dynamic hydrological prediction model is constructed by combining a deep long short-term memory network with a gated recurrent unit fusion architecture, including an input layer, a feature extraction layer, a parallel fusion feature processing layer, and an output layer. The input layer receives the multi-source fusion dataset and matches corresponding key hydrological factors according to climate type. This includes: determining the current climate type based on climate zoning data of the watershed area; constructing a mapping table between climate type and key hydrological factors based on historical hydrological data statistics and expert knowledge; recording key hydrological factors that affect runoff and water level changes under different climate types and their corresponding weights, such as rainfall and soil entropy under the rainy season climate type, and temperature and evaporation under the dry season climate type; automatically retrieving the mapping table according to the current climate type, selecting the corresponding key hydrological factors, and extracting the time series data of key hydrological factors from the multi-source fusion dataset; performing dimensional normalization and standardization on the selected key hydrological factor time series data, and outputting the normalized multidimensional feature vector; wherein, dimensional normalization is used to uniformly interpolate the key hydrological factor time series data of different time granularities to a 1-hour time granularity to ensure the synchronization of time series data; the standardization process uses the Z-score normalization method to map the key hydrological factor data to the [0,1] interval, effectively eliminating the differences in the dimensions of different factors. The feature extraction layer receives multidimensional feature vectors and groups them according to climate type. For features within each climate group, an improved Relief-F algorithm is used to calculate the dissimilarity among nearest neighbor samples. This includes: for samples within each climate group, filtering for similar and dissimilar nearest neighbors; calculating the dimensionality difference between each input feature and its nearest neighbors; normalizing the dimensionality difference by dividing it by the global range of each input feature; summing the squares of all normalized dimensionality differences and taking the square root to obtain the dissimilarity between each input feature and its nearest neighbors, including dissimilarity between dissimilar and similar nearest neighbors; calculating the difference between dissimilar and similar nearest neighbor dissimilarity, and then dividing this difference by the climate weights. The associated feature weights are obtained by multiplying the weights by the factors and adding them to the initial weights. A screening threshold is set based on historical hydrological data statistics and domain expert experience to dynamically screen key hydrological features. If the associated feature weight of the input feature is greater than the screening threshold, the input feature is determined to be a key hydrological feature and retained; otherwise, the input feature is determined to be a redundant feature and temporarily stored in the feature pool. For the retained key hydrological features, the corresponding time-series data are weighted using the associated feature weight as a coefficient, and the weighted key feature matrix is ​​output after being arranged according to time sequence. The climate weight factor is determined by the proportion of the influence intensity of different climate types on hydrological prediction indicators based on historical hydrological data statistics, and after calibration using domain expert experience. The parallel fusion feature processing layer receives the weighted key feature matrix. Based on a deep long short-term memory network and gated recurrent units, parallel processing units are constructed according to a first preset ratio, including a first processing unit and a second processing unit. The first processing unit slices the temporal features with a step size of 24, preserving long-term temporal correlation information. Through the input gate, forget gate, and output gate of the LSTM, the sliced ​​features are processed hourly. The forget gate automatically ignores short-term fluctuation noise. The input and output gates capture the changing trends of long-term hydrological data, such as seasonal runoff increase / decrease patterns. The output... The second processing unit uses time-series slices with a step size of 1 to enhance the capture of short-term time-series details. Through the update and reset gates of the GRU, it quickly responds to feature mutations caused by short-term sudden hydrological events, such as flow mutations caused by short-term rainstorms. It automatically focuses on key nodes with feature changes in a short period of time and outputs short-cycle feature vectors. Dual-path feature extraction is performed on the weighted key feature matrix. The long-cycle feature vector and the short-cycle feature vector are weighted and summed dimension by dimension to obtain the hydrological fusion feature vector. In this embodiment, the first preset ratio is set to 3:2. The output layer receives the hydrological fusion feature vector, maps it to the hydrological prediction result through a fully connected network, and outputs the preliminary confidence correlation value of the hydrological prediction result. Historical hydrological data sets are obtained through the interface between the historical hydrological database and the watershed monitoring system. The multi-source fusion dataset is then divided into training and validation sets according to a second preset ratio. The mean square error and Nash efficiency coefficient are used as the joint loss function to train the AI ​​dynamic hydrological prediction model using the Adam optimizer. After training, real-time multi-source fusion data is input into the AI ​​dynamic hydrological prediction model, and the initial hydrological prediction results and preliminary confidence scores are output. In this embodiment, the second preset ratio is set to 4:1.

[0024] For example, taking a watershed in southern China as an application scenario, after the input layer receives the multi-source fusion dataset, it determines that the current climate is the rainy season based on the watershed climate zoning data. It then retrieves the mapping table to select two key hydrological factors: rainfall and soil entropy. The initial weight of rainfall is 0.6 and the initial weight of soil entropy is 0.4. The time series data of rainfall and soil entropy from the previous month are extracted from the multi-source fusion dataset, and dimensionality normalization and standardization are performed to output the normalized multidimensional feature vector. The feature extraction layer receives multi-dimensional feature vectors and groups them according to the rainy season climate. For features within a group, an improved Relief-F algorithm is used to calculate the dissimilarity among nearest neighbor samples. This includes: selecting similar and dissimilar nearest neighbors from 1000 samples within the group; calculating the dimensionality difference between the current sample and its nearest neighbors for rainfall (0.12) and the soil entropy feature (0.09); dividing these two dimensionality differences by their respective global ranges to obtain normalized dimensionality differences of 0.14 and 0.10, respectively; summing the squares of the normalized dimensionality differences and taking the square root to obtain the dissimilarity between dissimilar nearest neighbors for rainfall (0.32) and similar nearest neighbors (0.18); and the dissimilarity between dissimilar nearest neighbors for soil entropy (0.25) and similar nearest neighbors (0.15). The rainfall... The difference between the dissimilar features and their nearest neighbors is 0.32 - 0.18 = 0.14. The difference between the dissimilar features and their nearest neighbors is 0.25 - 0.15 = 0.10. Multiplying the difference by the rainy season climate weight factor of 1.2 and adding it to the initial weight of 1.0, we get the weight of the rainfall-related feature as 1.0 + 1.2 × 0.14 = 1.168 and the weight of the soil entropy-related feature as 1.0 + 1.2 × 0.10 = 1.120. Setting the screening threshold to 0.2, if the weights of the two types of features are both greater than 0.2, they are identified as key hydrological features and retained. For the two retained key hydrological features, the corresponding time series data are weighted using their respective weights as coefficients, and the weighted key feature matrix is ​​output after being arranged according to the 720-hour time series. After receiving the weighted key feature matrix, the parallel fusion processing layer first processes the weighted key features into time-series slices with a step size of 24, preserving long-term temporal correlation information. The sliced ​​features are then processed hourly using the input, forget, and output gates of an LSTM, outputting a long-term feature vector. The second processing unit uses time-series slices with a step size of 1 to enhance short-term temporal detail capture. Through the update and reset gates of a GRU, it quickly responds to temporal abrupt changes caused by short-term heavy rainfall, automatically focusing on key nodes with large feature changes within a short time, such as those with normalized rainfall exceeding 0.8, and outputting a short-term feature vector. After completing dual-path feature extraction on the weighted key feature matrix, the long-term and short-term feature vectors are weighted and summed dimension-wise with a weight ratio of 0.6:0.4. For example, if the long-term vector has a dimension of 0.8 and the corresponding short-term vector has a dimension of 0.9, the sum is 0.8 × 0.6 + 0.9 × 0.4 = 0.84, yielding the hydrological fusion feature vector. After receiving the hydrological fusion feature vector, the output layer maps it to the initial hydrological prediction result through a 3-layer fully connected network and outputs the preliminary confidence correlation value. During the training phase of the AI ​​dynamic hydrological prediction model, a historical hourly hydrological measurement dataset of the watershed over the past 10 years was acquired. This dataset, combined with the aforementioned multi-source fusion dataset, was divided into training and validation sets at a pre-defined ratio of 4:1. The mean squared error (MSE) and Nash efficiency coefficient (NSE) were used as the joint loss function, with the formula Loss = 0.6 × MSE + 0.4 × (1 - NSE). The AI ​​dynamic hydrological prediction model was trained using the Adam optimizer. Early shutdown was triggered if the validation set error did not decrease for five consecutive rounds during training. The final validation set had MSE=0.025 and NSE=0.91. After training, the input watershed real-time multi-source fusion data were 320mm of granular rainfall and 72% of soil entropy in one hour of rainy weather. After standardization, the values ​​were 0.78 and 0.85, respectively, and were input into the AI ​​dynamic hydrological prediction model. The final output was the initial hydrological prediction result with a preliminary confidence level of 0.86. The initial hydrological prediction result was that the inflow rate in the reservoir in the next 24 hours would be 850m³ / s, the water level change would be 0.35m, and the peak runoff would be 1200m³ / s.

[0025] Preferably, the specific steps for outputting hydrological prediction results and hydrological confidence levels include: Historical hydrological data from the same period are used to compare the initial hydrological prediction results output by the AI ​​dynamic hydrological prediction model on a time-by-time basis. A hydrological error feedback function is constructed based on MSE and NSE. ;in, These are weighting coefficients, adjusted according to climate type, such as the rainy season. ,dry season , To account for comprehensive hydrological errors; Based on the error feedback function, a dynamic correction factor is set if Then the correction factor Take 1.0, no correction; if Then the correction factor Calculated using linear interpolation, the interpolation formula is as follows: ,like Time correction factor ;like This triggers a feature pool call, supplements non-critical factors, and re-predicts and corrects the results, yielding a corrected hydrological prediction. , Error distribution patterns were statistically analyzed based on historical forecast data from the past five years. The accuracy requirements for hydrological forecasting were determined based on calibration and the experience of experts in the field. pass The hydrological confidence level was calculated, and the confidence threshold was set based on historical prediction accuracy statistics and the experience of experts in the field of hydrology. If the hydrological confidence level is greater than the confidence threshold, it is judged as a high-confidence prediction result; otherwise, it is judged as a low-confidence prediction result. Real-time monitoring of corrected hydrological forecasts; setting extreme scenario trigger thresholds based on watershed hydrological disaster prevention and control standards and historical extreme event characteristics, including extreme inflow thresholds. Extreme threshold for water level change rate With peak runoff extreme threshold If the inflow in the revised hydrological forecast is greater than... Or the rate of change of water level is greater than Or peak runoff greater than This triggers the reinforcement learning branch of the AI ​​dynamic hydrological prediction model, which retrieves similar events from the historical case set based on the K-nearest neighbor algorithm. This includes: defining the retrieval feature dimensions; normalizing the feature vectors of the current extreme scenario with those of extreme events in the historical case set; calculating feature similarity using Euclidean distance; selecting the K cases with the smallest distance as similar events; extracting multi-source data and optimal hydrological prediction parameters from similar events as specialized training samples; and using the PPO reinforcement learning algorithm for specialized training. This includes: using the specialized training samples as training data; setting an agent responsible for adjusting parameters in the parallel fusion processing layer and output layer; designing a reward function based on comprehensive hydrological error; iteratively outputting parameter adjustment actions and providing feedback reward values ​​through the agent; limiting the parameter update amplitude and adjusting the learning rate based on the PPO algorithm; and performing training for three consecutive rounds. Training is stopped when the system is ready to output stable hydrological prediction results for extreme scenarios. These extreme scenarios include flood runoff caused by torrential rains and low water runoff caused by prolonged droughts. Designed based on the experience of experts in the field of hydrology; By integrating and correcting the hydrological forecast results for each 96-hour period and the hydrological forecast results for extreme scenarios, the continuity of results for each period is ensured. The accuracy of the hydrological forecast results is verified by obtaining the measured hydrological values ​​through the interface of the historical hydrological database and the watershed real-time monitoring system. The absolute difference between the hydrological forecast results and the measured hydrological values ​​is calculated. The absolute difference is divided by the measured hydrological values ​​to obtain the result error. According to the accuracy standards of the hydrological forecast industry and the requirements of engineering applications, the error threshold is set to 8%. For periods exceeding the error threshold, the hydrological error feedback function is used to correct the result until the result error is less than or equal to the error threshold. Finally, the high-precision hydrological forecast results and hydrological confidence scores for each period of the next 96 hours are output.

[0026] Example 3: This embodiment provides a detailed description of the collaborative scheduling module based on embodiment 1 or 2.

[0027] The collaborative scheduling module is used to construct a joint hydrological scheduling rule base based on hydrological forecast results. It combines hydrological confidence with hydrological change trends to dynamically optimize core scheduling parameters. High-confidence forecast results correspond to threshold control, while low-confidence forecast results reserve two levels of flexible adjustment space. A watershed hydraulic topology map is generated through a topology analysis algorithm, marking the upstream and downstream relationships of cascade reservoirs. Historical scheduling data is used to quantify the response time difference and impact of upstream reservoir scheduling actions on downstream water levels and flows, establishing a time difference tiered compensation rule to compensate for hydraulic transmission delays between upstream and downstream reservoirs. A distributed collaborative decision-making mechanism is introduced, establishing a encrypted real-time data sharing channel for cascade reservoirs. Each reservoir node synchronously updates its operating status and forecast data, dynamically allocating scheduling permissions and execution priorities based on demand priority. Core scheduling parameters include flood discharge flow adjustment weights, water storage level thresholds, and ecological benchmark protection standards. The collaborative scheduling module includes: a rule optimization unit, a time difference compensation unit, and a collaborative decision-making unit; The rule optimization unit receives the hydrological prediction results and hydrological confidence levels output by the AI ​​dynamic hydrological prediction model, constructs and dynamically updates the joint hydrological scheduling rule base based on hydrological change trends, quantifies and optimizes the adjustment weight of flood discharge flow, water storage level threshold, and ecological benchmark protection standards, and implements rigid threshold control for high-confidence prediction results, such as strictly controlling the water storage level to 0.5m below the warning level and rigidly allocating the flood discharge flow at 80% of the preset maximum safe flow; it configures two levels of flexible adjustment space for low-confidence prediction results, including primary adjustment and secondary adjustment, and finally generates the optimal scheduling rule set adapted to the current hydrological scenario. The time difference compensation unit is used to generate a watershed hydraulic topology map through a topology analysis algorithm, mark the upstream and downstream relationships of cascade reservoirs, retrieve historical scheduling data to quantify the response time difference and impact of upstream reservoir scheduling actions on downstream water level and flow, establish multi-level time difference hierarchical compensation rules to dynamically compensate for the hydraulic transmission delay of upstream and downstream reservoirs, and output a scheduling action correction scheme after time delay compensation. The collaborative decision-making unit is used to build a real-time data sharing channel for encrypted cascade reservoirs through encrypted transmission protocols and edge node deployment technology, and introduces a distributed collaborative decision-making mechanism to update the operating status and forecast data of each reservoir node. Based on the priority of flood control, water storage and ecological needs, it dynamically allocates the scheduling authority and execution priority of each reservoir node.

[0028] Preferably, the specific steps for constructing the joint hydrological scheduling rule base include: For the inflow in the hydrological forecast results, the difference in inflow between the current period and the previous period is calculated by obtaining the inflow in the previous period. Flow thresholds are set based on statistical analysis of the flow change rate during typical flood events in the basin. , ,like If so, it is determined that the inbound flow is on an upward trend; if If the inbound flow rate is stable, it can be determined that the inbound flow rate is stable; if If so, it can be determined that the inbound flow is showing a downward trend; For water level changes in hydrological forecasts, the difference between the current water level and the previous water level is calculated by obtaining the water level of the previous period. The water level threshold is set according to the reservoir water level monitoring accuracy, dam safety requirements, and downstream flood control standards. and ,like If so, it is determined that the water level change is trending upward; if If so, it is determined that the water level change is showing a stable trend; if If so, it can be determined that the water level is trending downwards; For the peak runoff in the hydrological forecast results, the runoff difference between the current time period and the previous time period is calculated by obtaining the previous peak runoff. Runoff thresholds are set based on historical runoff peak variation patterns and watershed confluence characteristics. , ,like If so, it is determined that the peak runoff is trending upward; if If so, it is determined that the peak runoff shows a stable trend; if If so, it can be determined that the peak runoff is showing a downward trend; A weighted voting method was used, with inflow weighted at 0.4, water level change at 0.3, and peak runoff at 0.3. The trend voting results of the three indicators were statistically analyzed. If the product of the number of rising votes and the corresponding weight was the largest, the overall hydrological trend was determined to be rising; if the product of the number of stable votes and the corresponding weight was the largest, the overall hydrological trend was determined to be stable; and if the product of the number of falling votes and the corresponding weight was the largest, the overall hydrological trend was determined to be falling. Set the trend coefficient of hydrological change trend according to the type of hydrological change trend. Including the upward trend coefficient The stability trend coefficient is and the downward trend coefficient is Initial scheduling parameters, including the baseline value of flood discharge flow weight, are obtained from a known scheduling parameter library. Initial threshold of water storage level Initial values ​​for ecological baseline protection ; calculate and The product of these factors yields the adjustment weight for the flood discharge flow. In this embodiment, the flow influence coefficient is set to 0.1 and the flow base coefficient is set to 1. The flow correction coefficient is obtained by subtracting the product of the flow influence coefficient and the inflow from the flow base coefficient. The flow correction coefficient is then compared with the initial threshold of the water storage level. Multiply to obtain the water storage level threshold In this embodiment, the water level influence coefficient is set to 0.05 and the water level baseline coefficient is set to 1. The product of the water level influence coefficient and the water level change value is calculated and summed with the water level baseline coefficient to obtain the water level correction coefficient. The water level correction coefficient is then multiplied by the initial value of the ecological benchmark protection. Obtain ecological benchmark protection value ; For high-confidence predictions, rigid threshold controls are implemented, such as strictly controlling the water level to 0.5m below the warning level and rigidly allocating the flood discharge flow at 80% of the preset maximum safe flow. For low-confidence predictions, two levels of flexible adjustment space are configured, including primary and secondary adjustments. When the hydrological confidence level is 75%-85%, primary adjustment is performed, with the water level finely adjusted within ±0.2m of the baseline threshold and the flood discharge flow adjustment not exceeding 10%. When the hydrological confidence level is 60%-74%, secondary adjustment is performed, with the water level adjusted within ±0.4m of the baseline threshold and the flood discharge flow adjustment not exceeding 20%, and an emergency adjustment interface is reserved. Finally, an optimal scheduling rule set adapted to the current hydrological scenario is generated. The preset maximum safe flow is set based on the design flood discharge capacity of the cascade reservoir dams in the basin, the maximum flood carrying capacity of the river channel, and the safety standards of the downstream flood protection zone. The optimal scheduling rule set is classified and stored according to trend type, confidence level and parameter range, and integrated with historical rules to form a structured, queryable and updatable rule system, thereby constructing a hydrological joint scheduling rule base.

[0029] Preferably, the specific steps for compensating for the hydraulic transmission time delay in upstream and downstream reservoirs include: The geographical coordinates of each reservoir and the upstream and downstream river connections are transformed using a topology analysis algorithm. As a directed graph structure, with reservoir nodes as vertices and upstream and downstream river connections as directed edges, a watershed hydraulic topology graph is generated, labeling the upstream and downstream relationships of cascade reservoirs, the ID of each reservoir node, and the river length; Extracting the first time point of a single flood discharge / water storage action from upstream reservoirs from historical scheduling data, and matching it with the second time point where the changes in water level and flow rate of downstream reservoirs exceed a preset threshold, the difference between the first and second time points is calculated to obtain the hydraulic response time difference under a single scenario. In this embodiment, the first time point is set as the time when the action begins to be executed, and the second time point is the time when the change begins to have a significant response. The preset amplitude threshold includes the flow adjustment amplitude. Water level adjustment range ; Historical data were grouped according to hydrological trend types. The correlation coefficient between upstream scheduling adjustments and downstream water level / flow changes was calculated for each scenario to determine the impact correlation. The distribution of hydraulic response time differences for each scenario was statistically analyzed, and the 25th and 75th percentiles were selected to divide the time difference intervals. If, then it is divided into the first time difference interval; if If so, it is divided into the second time difference interval; if If the time difference interval is divided into the third time difference interval, the mean of the downstream change corresponding to the unit scheduling action adjustment amount in each group of scenarios is calculated as the quantitative value of the impact. For example, taking a cascade of reservoirs in a river basin, upstream reservoir A and downstream reservoir B are selected from three typical scenarios: hydrological upward trend, hydrological stable trend, and hydrological downward trend. Based on historical scheduling data from the past five years, after grouping according to hydrological trend type, data on upstream scheduling adjustments, downstream water level / flow changes, and hydraulic response time differences are extracted. Among these, the upward trend scenario... , , Stable trend scenario , , In a downward trend scenario, the main action is water storage. Negative values ​​are used; calculate the Pearson correlation coefficient between the upstream scheduling action adjustment and the downstream change for each scenario group, for scenarios with rising, stable, and falling trends. and The correlation coefficients were 0.92, 0.83, and 0.88, respectively. and The correlation coefficients were 0.95, 0.87, and 0.91, respectively, all indicating a strong correlation and confirming a significant impact. The hydraulic response time difference distribution for each scenario was statistically analyzed, and the 25th and 75th percentiles were calculated to divide the time difference intervals, such as for the upward trend scenario. , This is divided into three time zone intervals: the first, the second, and the third; (Stable trend scenario) , This is divided into three time zone intervals: the first, the second, and the third; (Downtrend scenario) , The time difference is divided into three intervals: the first, the second, and the third. The magnitude of the impact is quantified based on the average downstream change corresponding to a unit upstream scheduling action adjustment, including: the magnitude of the impact. Among them, the quantitative value of the impact of water level on the upward trend scenario. Quantitative value of the impact of traffic flow Quantitative value of the impact of water level on a stable trend scenario Quantitative value of the impact of traffic flow Ultimately, the impact magnitude of each scenario is quantified, clarifying the average impact of upstream scheduling actions on downstream water levels and flow rates. Based on the hydraulic response time difference interval, the quantified value of downstream impact, and the watershed hydraulic transmission delay control requirements, a time difference-based compensation rule is established using core parameters from the hydrological joint scheduling rule base to compensate for the hydraulic transmission delay of upstream and downstream reservoirs. This includes primary, secondary, and tertiary compensation. In this case, a first-level compensation is implemented, and the upstream reservoir synchronously sends scheduling actions to the downstream reservoir. The downstream reservoir then fine-tunes its water level based on real-time data, such as the adjustment range. To ensure timely and prompt response; if In this case, secondary compensation will be implemented, with upstream reservoirs receiving compensation in advance. Anticipate the impact of scheduling and adjust the weighting of flood discharge flow, such as the adjustment range. ;like In this case, a three-tiered compensation system will be implemented. Upstream reservoirs will execute phased scheduling actions, transmitting the anticipated impact and scheduling plan to downstream reservoirs one hour in advance. Downstream reservoirs will then adjust their water level thresholds in advance based on the magnitude of the impact. For example, if the adjustment range... Among them, the core parameters include, but are not limited to, the benchmark value for adjusting the flood discharge flow, the benchmark threshold for water storage level, and the ecological benchmark protection standard; A distributed data sharing channel is established using the national cryptographic SM2 / SM4 encryption protocol. Edge encryption gateways are deployed at each reservoir node to ensure the security of data transmission and storage. Each reservoir node synchronizes its own operational status data and global hydrological prediction results in real time through the distributed data sharing channel. The cloud coordinates the aggregation of data from all nodes to generate a global hydrological and operational status interface. Scheduling permissions are dynamically allocated based on demand priority, including: for flood control scenarios, upstream reservoir nodes have the highest scheduling permissions; for ecological water supply scenarios, downstream reservoir nodes have priority scheduling permissions; for power generation scenarios, permissions are dynamically allocated according to the efficiency of the reservoir source and the generating units, such as the efficiency of the headwater reservoir and the generating units. The reservoir is assigned priority scheduling authority, and the water level is adjusted according to the inflow forecast to ensure power generation efficiency while taking into account the flood control and ecological needs of the upstream and downstream areas; in this embodiment, the priority of needs is set as flood control > ecology > power generation.

[0030] Example 4: This embodiment provides a detailed description of the scheduling correction module based on any one of embodiments 1-3.

[0031] The scheduling correction module generates a set of scheduling instructions for the cascade reservoirs in the basin based on hydrological forecast results, the joint hydrological scheduling rule base, and time difference compensation rules. This set includes a first scheduling instruction and a second scheduling instruction. The first scheduling instruction is used for allocating storage and release volumes and adjusting power generation output in each reservoir under normal hydrological scenarios, determining the target water level and baseline value for release flow for each time period. The second scheduling instruction is used for emergency adjustments under extreme hydrological scenarios. It acquires real-time execution feedback data of the scheduling instruction set, calculates the comprehensive scheduling deviation by combining hydrological forecast results and scheduling target values, determines the deviation level, triggers graded responses, and dynamically corrects differentiated scheduling instructions to ensure scheduling coordination. In this embodiment, all scenarios except extreme hydrological scenarios are set as normal hydrological scenarios.

[0032] Preferably, the specific steps for determining the deviation level include: For conventional hydrological scenarios, based on the joint hydrological scheduling rule base, the weights of water storage and discharge and power generation are dynamically allocated, including a flood control weight of 0.4, a water storage weight of 0.3, and a power generation weight of 0.3. Based on the current water level, predicted inflow, and remaining reservoir capacity, the target water level for each reservoir in each time period is calculated. For example, if the current water level of Reservoir A is... Querying the storage capacity curve yields the current storage capacity, which will then be used to predict the inflow rate. Subtract the allowable discharge flow Obtain net inflow rate The net inflow rate is calculated by dividing the net inflow rate by the time period length of 30 minutes to obtain the net inflow volume for this period. Multiplying the net inflow during this period by the water storage weight of 0.3 yields the target water storage increment. The target reservoir capacity is obtained by summing the target water storage increment with the current reservoir capacity. The theoretical target water level is then obtained by looking up the reservoir capacity curve. Within the fine-tuning range allowed by the joint hydrological dispatching rules, such as ±0.2m, the theoretical target water level will be adjusted to the optimal target water level that balances power generation and ecological needs. The reservoir capacity curve is generated by fitting measured topographic data, underwater topographic scanning results, and historical water level and reservoir capacity records. The time-difference compensation correction method is used to determine the baseline value of the discharge flow. For example, based on the basin's flood control requirements, ecological base flow demand, and reservoir operation experience, the baseline discharge ratio coefficient under conventional hydrological scenarios is determined to be 0.6. The time-difference compensation coefficient is calculated as 1.3 based on the response time difference of upstream reservoir scheduling actions to downstream water level and flow. The time-difference compensation adjustment coefficient is set to 0.1 based on the downstream river's hydraulic conduction characteristics, historical response data, and scheduling safety margin. The baseline discharge flow is obtained by multiplying the predicted inflow flow by the baseline discharge ratio coefficient of 0.6. The time-difference compensation adjustment amount is obtained by multiplying the time-difference compensation coefficient by the time-difference compensation adjustment coefficient of 0.1. The baseline discharge flow of reservoir A is obtained by summing the baseline discharge flow and the time-difference compensation adjustment amount. Based on the output curve and water level target, such as the water level of reservoir B... The corresponding output is The target output value corresponding to the current water level is determined by linear interpolation, including: calculating the difference between the current water level and the lower limit of the interval based on the current water level interval; dividing the difference by the difference between the upper and lower limits of the interval to obtain the water level ratio coefficient; calculating the product of the water level ratio coefficient and the difference in the output interval; and summing the product with the lower limit of the output interval to obtain the target output value corresponding to the current water level. Finally, the first dispatch instruction is formed by integrating the current hydrological forecast results, hydrological change trend coefficient, hydrological confidence level, hydrological collaborative dispatch rule base and time difference compensation rules, including the time-period water level target, the downstream flow baseline value, and the output adjustment interval. The output curve is generated by fitting measured head, flow, output data, unit efficiency and historical operation records. In response to extreme hydrological scenarios, emergency dispatch rules are triggered, increasing the flood control priority weight to 0.8, the water storage weight to 0.1, and the power generation weight to 0.1. Emergency adjustment parameters from similar historical extreme scenarios are extracted in real time to determine the upper limit of the emergency water storage level. For example, if the upper limit of the water level in Reservoir A drops to [a certain value] in an extreme hydrological scenario... Lower than conventional hydrological scenarios Determine the upper limit of the emergency discharge flow, such as the upper limit of the emergency discharge for reservoir A being [missing information]. The emergency reduction range of power generation output is determined, such as reducing output by 30% to ensure flood discharge; a dynamic emergency adjustment interface is set up to support fine-tuning every 30 minutes based on real-time data, forming a second dispatch instruction, including the emergency water level / flow limit, output reduction requirements and emergency adjustment trigger conditions; The first and second scheduling instructions are integrated by timestamp to verify the coordination of cross-reservoir scheduling instructions. For example, does adjusting the discharge flow of Reservoir A cause the water level of Reservoir B to exceed the standard? If, after the basin hydraulic simulation shows that Reservoir A adjusts its discharge flow according to the integrated scheduling instructions, the actual water level of Reservoir B will exceed the water level control threshold of the corresponding scenario, then a conflict is determined, and correction is made based on the time difference compensation rule; otherwise, the coordination of instructions is determined to meet the standard, the scheduling actions of each reservoir do not conflict with each other, and the requirements of upstream and downstream hydraulic transmission and time difference compensation are met, and finally, a set of scheduling instructions for cascade reservoirs in the basin is generated. Among them, the water level control threshold includes the target water level ±0.2m under normal hydrological scenarios and the upper limit of emergency water level under extreme hydrological scenarios. Real-time acquisition of execution feedback data from the scheduling instruction set, including actual water level. Actual discharge flow With actual power generation For each time period and each reservoir, calculate the actual water level and the target water level. The absolute difference is used to obtain the water level deviation. Calculate the actual discharge flow rate and the baseline discharge flow rate. The absolute difference is used to obtain the flow deviation. Calculate the actual power generation output and the power generation adjustment benchmark value. The absolute difference is the output deviation. The power output regulation benchmark value is determined by weighted calculation based on the power output target value corresponding to the current water level, the rated output of the generating units, the ecological flow guarantee requirements, and the real-time grid load demand. Based on the core objectives of basin scheduling and historical deviation control experience, water level weight is set to 0.4, flow weight to 0.4, and power output weight to 0.2. The weighted deviation is obtained by multiplying the water level deviation by its weight, the flow deviation by its weight, and the power output deviation by its weight. The weighted deviation is then summed by multiplying the water level target by its weight, the downstream flow benchmark value by its weight, and the power output regulation benchmark value by its weight. The overall scheduling deviation is obtained by calculating the ratio of the weighted deviation to the benchmark weighted deviation. The expression is as follows:

[0033] The deviation threshold is set based on scheduling accuracy requirements, safety control standards for various scenarios, historical deviation response effects, and watershed hydraulic characteristics. , ,and Combined with comprehensive scheduling deviation Determine the deviation level; if If it is, then it is judged as a level one deviation; if If it is, then it is judged as a second-order deviation. If the deviation is not met, it is determined to be a level three deviation; in this embodiment, it is set as follows: , ; Based on the deviation level and scenario type, a tiered response is triggered. For routine hydrological scenarios, a Level 1 deviation triggers a passive response, recording only the deviation data without adjusting the first dispatch command; a Level 2 deviation triggers a dominant response, initiating the routine correction process; a Level 3 deviation triggers an emergency response, temporarily activating the emergency interface for the second dispatch command. For extreme hydrological scenarios, a Level 1 deviation triggers an enhanced monitoring response, shortening the feedback cycle to 5 minutes; a Level 2 deviation triggers an emergency fine-tuning response, setting the adjustment range based on the magnitude of the comprehensive dispatch deviation, the quantified value of the downstream impact, and the safety boundary of the emergency parameters, adjusting the emergency parameters accordingly. These emergency parameters include the upper limit of the emergency water storage level, the upper limit of the emergency discharge flow, and the emergency reduction ratio of power generation output; a Level 3 deviation triggers the highest-level emergency response, immediately suspending power generation output to prioritize flood discharge and water level safety. Due to the uncertainties in hydrological forecasting, the dynamic changes in real-time operating status, and the time delays in upstream and downstream hydraulic transmission, the initial dispatch instructions may not be fully adapted to the actual scenario. Therefore, based on deviation feedback and collaborative management requirements, differentiated dispatch instructions need to be dynamically corrected. This includes: for secondary deviations in conventional hydrological scenarios, a fine-tuning method using index weights is used for correction; for tertiary deviations, the emergency water level upper limit of the second dispatch instruction is temporarily invoked, the target water level is lowered by 0.2m, the baseline value of the discharge flow is increased by 20%, and the same applies to downstream reservoirs; for secondary deviations in extreme hydrological scenarios, emergency parameters are corrected according to the quantified value of the impact magnitude; for tertiary deviations, a shutdown and discharge protection instruction is immediately generated, all generator units are suspended, the discharge flow is increased to the emergency upper limit, and upstream and downstream collaborative compensation is initiated. For example, in a typical hydrological scenario during time period a, the initial scheduling command parameters for reservoir A are: target water level. Downflow baseline value Output adjustment reference value The weights for flood control and power generation are 0.4 for flood control, 0.3 for water storage, and 0.3 for power generation; the initial scheduling command parameters for Reservoir B are: target water level. Downflow baseline value Output adjustment reference value The weight configuration is consistent with that of Reservoir A; real-time execution feedback data of the scheduling instruction set is obtained for Reservoir A, including the actual water level. The actual discharge flow rate is The actual power output is Water level deviation Flow deviation , output deviation The weighted total deviation is The benchmark weighted sum is Comprehensive scheduling deviation , In and The deviation was determined to be a Level 2 deviation. For Level 2 deviations, a fine-tuning method of indicator weights was used for correction. This included: based on the principle of prioritizing safety while considering power generation and ecological considerations, the water storage weight was reduced from 0.3 to 0.25, with the 0.05 reduction allocated to the flood control weight; the water storage weight was increased from 0.4 to 0.45, while the power generation weight remained unchanged at 0.3. The corrected scheduling parameters were calculated based on the new weights. The corrected target water storage increment was the product of the net inflow during the current period and the new water storage weights. The corrected target reservoir capacity is the current reservoir capacity plus the corrected target water storage increment. Combining this with the contrast of the reservoir capacity curve, the corrected water level target is obtained. The revised baseline value of the outflow = (predicted inflow × basic outflow ratio coefficient) + (time difference compensation coefficient + time difference compensation adjustment coefficient) × new flood control weight correction coefficient = The new flood control weight correction factor is 0.45 / 0.4 = 1.125; based on the corrected water level. Based on the power output curve of reservoir A, The corresponding output is 60-70. The water level proportionality coefficient is calculated using linear interpolation. The revised output target is Concurrently, due to a lag in the initial discharge adjustment of upstream reservoir A, reservoir B experienced an overall scheduling deviation. , Greater than The deviation was determined to be Level 3, and the emergency water level cap of the second dispatch order was immediately activated to raise the original target water level. Downgraded to Original discharge baseline value Increase by 20% The data is then synchronized to downstream reservoir C via an encrypted data sharing channel, notifying reservoir C 0.5 hours in advance to prepare for pre-release, compensating for hydraulic transmission delays, and ensuring the safety of flood discharge in the downstream river channel; among which, This refers to the actual water level of reservoir B in real time during period a. This refers to the actual discharge flow rate. This refers to the actual power generation output.

[0034] In summary, the present invention discloses an optimization method for joint scheduling of cascade reservoirs in a watershed based on AI hydrological prediction. This system aims to address the problems of slow joint scheduling response and low coordination efficiency caused by low prediction accuracy and the lack of dynamic collaborative scheduling mechanisms among cascade reservoirs in a watershed. It achieves high-precision, collaborative, and dynamic scheduling of cascade reservoirs through four modules: data fusion, hydrological prediction, collaborative scheduling, and scheduling correction. The data fusion module acquires multi-source data through the Internet of Things, encrypted transmission, and distributed web crawling. After denoising, outlier removal, time-series synchronization, and feature alignment, it generates a standardized multi-source fusion dataset, laying the data foundation for subsequent prediction and scheduling. The hydrological prediction module constructs an AI dynamic hydrological prediction model that integrates LSTM and GRU, embedding adaptive feature extraction and incremental learning mechanisms. It automatically matches climate types and key hydrological factors, captures long-term and short-term hydrological patterns through dual-path feature processing, and trains and optimizes the AI ​​dynamic hydrological prediction model by combining a joint loss function of mean square error and Nash efficiency coefficient. A confidence assessment and dynamic correction mechanism is introduced, triggering reinforcement learning branches in extreme hydrological scenarios, ultimately outputting high-precision hydrological prediction results and hydrological confidence levels for each time period of the next 96 hours. The collaborative scheduling module constructs a dynamic joint scheduling rule base based on hydrological prediction results and hydrological confidence levels. It optimizes scheduling parameters by combining hydrological change trends, implementing rigid threshold control for high-confidence prediction results and reserving two levels of flexible adjustment space for low-confidence prediction results. It generates a watershed hydraulic topology map through topology analysis, quantifies the response time difference and impact magnitude between upstream and downstream areas, and establishes hierarchical time difference compensation rules to compensate for hydraulic transmission delays in upstream and downstream reservoirs. A distributed collaborative decision-making mechanism is adopted, allocating scheduling permissions according to the priority order of flood control > ecology > power generation to ensure coordinated response between upstream and downstream areas. The scheduling correction module generates first scheduling instructions for conventional hydrological scenarios and second scheduling instructions for extreme water level scenarios, integrating and verifying coordination by timestamp to avoid cross-reservoir scheduling conflicts. It collects instruction execution feedback data in real time, calculates three types of deviations: water level, flow rate, and output, obtains a comprehensive scheduling deviation through weighted calculation, and classifies the deviation. Differentiated correction responses are triggered for different deviation levels and scenario types, dynamically adjusting scheduling parameters and weights to ensure that scheduling instructions adapt to actual hydrological changes and hydraulic transmission characteristics, comprehensively guaranteeing the balance of multiple objectives: flood control, ecology, and power generation in the watershed.

[0035] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A watershed cascade reservoir joint scheduling optimization system based on AI hydrological prediction, characterized in that, include: The data fusion module is used to acquire real-time monitoring data, real-time meteorological data and historical hydrological data in the watershed. After preprocessing, it generates a multi-source fusion dataset and outputs it to the hydrological prediction module. The hydrological prediction module is used to receive the multi-source fusion dataset, construct an AI dynamic hydrological prediction model, automatically identify the weights of key hydrological factors under different climatic conditions, iteratively optimize model parameters using an incremental learning mechanism, introduce a confidence assessment mechanism to construct a hydrological error feedback function to dynamically correct prediction bias, trigger reinforcement learning branches for special training in extreme hydrological scenarios, and output hydrological prediction results and hydrological confidence to the collaborative scheduling module. The collaborative scheduling module receives the hydrological prediction results and hydrological confidence levels, constructs and dynamically updates the joint hydrological scheduling rule base, optimizes the core scheduling parameters by combining hydrological confidence levels and hydrological change trends, and configures the control mode according to the confidence level. It generates a watershed hydraulic topology map through topology analysis, quantifies the response time difference and impact magnitude of upstream and downstream scheduling, establishes time difference hierarchical compensation rules to compensate for hydraulic transmission delays, and introduces a distributed collaborative decision-making mechanism to achieve real-time data sharing and dynamic allocation of scheduling permissions. It outputs the optimal scheduling rule set, time difference compensation scheme, and scheduling permission allocation results to the scheduling correction module. The scheduling correction module is used to receive the optimal scheduling rule set, time difference compensation scheme and scheduling authority allocation results, generate a first scheduling instruction adapted to conventional hydrological scenarios and a second scheduling instruction adapted to extreme hydrological scenarios, and integrate them to form a basin-wide cascade reservoir scheduling instruction set; obtain scheduling execution feedback data in real time to calculate comprehensive scheduling deviation, determine the deviation level and trigger graded responses, and dynamically correct differentiated scheduling instructions.

2. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction as described in claim 1, characterized in that, The preprocessing process of the data fusion module is as follows: A wavelet threshold denoising algorithm is used to filter electromagnetic interference noise in real-time monitoring data, abnormal extreme value data is removed based on the 3σ criterion, and time sequence alignment of multi-site data is achieved through a timestamp synchronization algorithm. A feature alignment and fusion algorithm is used to associate and preprocessed real-time monitoring data, real-time meteorological data, and historical hydrological data to generate a standardized multi-source fusion dataset. The multi-source fusion dataset includes instantaneous rainfall intensity, real-time reservoir water level, inflow / outflow, meteorological forecast data, historical runoff data, and static parameters of cascade reservoirs.

3. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction as described in claim 1, characterized in that, The specific steps for the hydrological prediction module to construct the AI ​​dynamic hydrological prediction model are as follows: Based on the multi-source fusion dataset, an AI dynamic hydrological prediction model is constructed, which includes an input layer, a feature extraction layer, a parallel fusion feature processing layer, and an output layer. The input layer receives a multi-source fusion dataset, matches the corresponding key hydrological factors according to the climate type, extracts time-series data, and outputs a multi-dimensional feature vector. The feature extraction layer receives the multidimensional feature vector and groups them according to climate type. It uses the improved Relief-F algorithm to calculate the difference between the dissimilarity of each feature in the nearest neighbor samples and the dissimilarity of the same type of nearest neighbor. It then calculates the associated feature weights by combining the climate weight factor with the initial weights. If the weight of the associated feature is greater than the preset screening threshold, it is determined to be a key hydrological feature and is weighted to output a weighted key feature matrix. Otherwise, it is identified as a redundant feature and temporarily stored in the feature pool; The parallel fusion feature processing layer constructs parallel processing units according to the first preset ratio. The first processing unit extracts long-period feature vectors from time slices with a step size of 24 using LSTM, and the second processing unit extracts short-period feature vectors from time slices with a step size of 1 using GRU. After performing dual-path feature extraction on the weighted key feature matrix, the weighted sum is obtained to obtain the hydrological fusion feature vector. The output layer receives the hydrological fusion feature vector, maps it to the initial hydrological prediction result through a fully connected network, and outputs the preliminary confidence correlation value. Historical hydrological measurement datasets are obtained, and the multi-source fusion dataset is divided into training and validation sets according to a second preset ratio. A joint loss function and Adam optimizer are used to train an AI dynamic hydrological prediction model.

4. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to claim 3, characterized in that, The specific steps for the hydrological prediction module to output hydrological prediction results and hydrological confidence levels are as follows: By calling up historical hydrological data from the same period and comparing them with the initial hydrological prediction results, a hydrological error feedback function based on mean square error and Nash efficiency coefficient is constructed, and a dynamic correction factor is set to obtain the corrected hydrological prediction results. The hydrological confidence level is calculated by subtracting the comprehensive hydrological error from 10, and the high-confidence prediction results are determined by combining the pre-set confidence threshold. Set extreme scenario trigger thresholds θ 1. θ 2. θ 3. If the inflow in the revised hydrological forecast is greater than... θ 1 or the rate of water level change is greater than θ 2 or peak runoff greater than θ 3. Then the reinforcement learning branch is triggered, which retrieves multi-source data of similar events and the optimal hydrological prediction parameters as special training samples for training, and outputs the hydrological prediction results for extreme scenarios. By integrating the corrected hydrological prediction results with those for extreme scenarios, the error in the calculation of measured hydrological values ​​is obtained. For time periods exceeding the preset error threshold, a hydrological error feedback function is used for re-correction. Finally, the hydrological prediction results and hydrological confidence levels for each time period in the next 96 hours are output.

5. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction as described in claim 1, characterized in that, The collaborative scheduling module includes a rule optimization unit, a time difference compensation unit, and a collaborative decision-making unit: The rule optimization unit is used to receive the hydrological prediction results and hydrological confidence levels, construct and dynamically update the hydrological joint scheduling rule base in combination with hydrological change trends, quantify the core scheduling parameters, configure differentiated control modes for prediction results with different confidence levels, and generate the optimal scheduling rule set. The time difference compensation unit is used to generate a watershed hydraulic topology map and mark the upstream and downstream relationships of cascade reservoirs, retrieve historical scheduling data to quantify the response time difference and impact of upstream and downstream scheduling, establish multi-level time difference hierarchical compensation rules to compensate for hydraulic transmission delay, and output a time difference compensation scheme. The collaborative decision-making unit is used to establish a real-time data sharing channel for encrypted cascade reservoirs, introduce a distributed collaborative decision-making mechanism to update the operating status and forecast data of each reservoir node, dynamically allocate scheduling permissions and execution priorities based on the demand priority of flood control > ecology > power generation, and output the scheduling permission allocation results.

6. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to claim 5, characterized in that, The specific steps for the rule optimization unit to construct the hydrological joint scheduling rule base are as follows: For the three types of indicators in the hydrological forecast results—inflow, water level change, and peak runoff—the difference between the current period and the previous period is calculated, and the trend of each indicator is determined based on the corresponding threshold. The three types of indicators are assigned corresponding weights. Based on the product of the trend voting results of each indicator and its corresponding weight, the overall hydrological trend is determined to be rising, stable, or falling. Based on the hydrological change trend, a trend coefficient is set, and the corrected water storage level threshold, ecological benchmark guarantee value and flood discharge flow adjustment weight are calculated in combination with the initial scheduling parameters. A rigid threshold control is applied to high-confidence prediction results, while a flexible space for first-level and second-level adjustments is configured for low-confidence prediction results, thereby generating an optimal scheduling rule set. The optimal scheduling rule set is stored according to trend type, confidence level and parameter range, and historical rules are integrated to build a hydrological joint scheduling rule base.

7. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to claim 5, characterized in that, The specific steps for the time difference compensation unit to quantify the time difference and impact of upstream and downstream scheduling responses are as follows: The geographical coordinates of each reservoir and the connection relationship between upstream and downstream river channels are transformed into a directed graph structure, with reservoir nodes as vertices and upstream and downstream river channel connections as directed edges, to generate a watershed hydraulic topology graph. Extract the first time point of a single flood discharge / water storage action of the upstream reservoir, match it with the second time point where the water level / flow rate change of the downstream reservoir exceeds a preset threshold, and calculate the difference to obtain the hydraulic response time difference under a single scenario. C ; Historical data were grouped according to hydrological trend type, and the correlation coefficient between the upstream scheduling action adjustment and the downstream water level / flow change was calculated for each group of scenarios. The distribution of hydraulic response time difference in each scenario was statistically analyzed. The 25th and 75th percentiles were selected to divide the time difference into three intervals. The mean downstream change corresponding to the unit scheduling action adjustment in each scenario was calculated as the quantitative value of the impact. Establish time zone compensation rules: If C If the time is ≤1 hour, Level 1 compensation will be applied; if the time is <1 hour C If the time is ≤3 hours, a secondary compensation will be applied. C If the time exceeds 3 hours, a three-level compensation will be applied.

8. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to claim 1, characterized in that, The specific steps for the scheduling correction module to generate the scheduling instruction set are as follows: For conventional hydrological scenarios, the weights of flood control, water storage, and power generation are dynamically allocated based on the optimal scheduling rule base. The target water level for each reservoir during a given time period is calculated by combining the current water level, the predicted inflow, and the reservoir capacity. The baseline value of the outflow is determined based on the basic outflow ratio coefficient, the response time difference compensation coefficient, and the adjustment coefficient. The target power generation value is calculated by interpolation based on the target water level and the power output curve, thus forming the first scheduling instruction. In response to extreme hydrological scenarios, emergency dispatch rules are triggered and flood control priority is increased. Emergency parameters from similar historical cases are extracted to determine the upper limit of emergency water level and flow rate and the range of power generation output reduction. A dynamic emergency adjustment interface is set up to generate a second dispatch instruction. The first and second scheduling instructions are integrated according to timestamps, the coordination of cross-reservoir scheduling instructions is verified, and conflicting instructions are corrected based on the time difference compensation scheme to generate a set of scheduling instructions for cascade reservoirs in the basin.

9. The basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to claim 8, characterized in that, The specific steps for the scheduling correction module to determine the deviation level and make dynamic corrections are as follows: Real-time acquisition of execution feedback data of scheduling instruction set, and calculation of the absolute deviation between the actual values ​​of three types of indicators—water level, flow rate, and output—and the benchmark target values ​​for each time period and each reservoir; Assign corresponding weights to the three types of deviations, and calculate the ratio of the weighted sum of deviations to the baseline weighted sum to obtain the comprehensive scheduling deviation ΔG; Set deviation thresholds γ1 and γ2. If ΔG≤γ1, it is judged as a first-level deviation; if γ1<ΔG≤γ2, it is judged as a second-level deviation; if ΔG>γ2, it is judged as a third-level deviation. Based on the deviation level and scenario type, a graded response is triggered: In a normal hydrological scenario, a first-level deviation only records data, a second-level deviation initiates a normal correction process, and a third-level deviation temporarily activates the second dispatch command emergency interface; In an extreme hydrological scenario, a first-level deviation shortens the feedback cycle, a second-level deviation fine-tunes emergency parameters, and a third-level deviation immediately triggers a shutdown and release command. The scheduling parameters are dynamically adjusted based on the correction strategy, and the corrected instructions are synchronized to all cascade reservoir nodes.

10. The joint scheduling optimization method for a basin-wide cascade reservoir joint scheduling optimization system based on AI hydrological prediction according to any one of claims 1-9, characterized in that, Includes the following steps: S1. Data Fusion: Acquire multi-source hydrological and meteorological data within the watershed, and generate a multi-source fused dataset after preprocessing; S2, AI hydrological prediction: Based on the multi-source fusion dataset, an AI dynamic hydrological prediction model is constructed. Through adaptive feature extraction, dual-path time series modeling and dynamic error correction, the hydrological prediction results and hydrological confidence are output. S3. Collaborative Scheduling: Based on the hydrological prediction results and hydrological confidence level, a dynamic hydrological joint scheduling rule base is constructed, the upstream and downstream hydraulic response time difference is quantified and a hierarchical compensation mechanism is established, and scheduling permissions are allocated through distributed collaborative decision-making. S4. Scheduling Correction and Execution: Generate differentiated scheduling instruction sets for normal and extreme scenarios, monitor execution feedback in real time and calculate comprehensive scheduling deviations, trigger hierarchical response to dynamically correct scheduling instructions, and realize joint optimization scheduling of cascade reservoirs in the basin.

11. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the joint scheduling optimization method of the basin cascade reservoir joint scheduling optimization system based on AI hydrological prediction as described in claim 10.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the joint scheduling optimization method of the basin cascade reservoir joint scheduling optimization system based on AI hydrological prediction as described in claim 10.

Images