Intelligent reservoir multi-objective scheduling method and system based on artificial intelligence

Abstract

This invention relates to the field of intelligent scheduling technology for water conservancy projects, specifically to a smart reservoir multi-objective scheduling method and system based on artificial intelligence. The method includes: collecting multi-dimensional historical operational data of the target reservoir; extracting spatiotemporal features to generate a reservoir system feature vector containing periodic components of runoff, water level fluctuation modes, outflow correlation coefficients, power generation load characteristics, and ecological water demand satisfaction indicators; constructing a deep neural network multi-objective optimization model incorporating runoff prediction, power generation optimization, and ecological scheduling sub-networks; obtaining inflow runoff prediction values ​​through the runoff prediction sub-network; generating an initial scheduling scheme through the power generation optimization sub-network; and finally, outputting a comprehensive scheduling scheme that coordinates power generation and ecological objectives through the ecological scheduling sub-network. This method can comprehensively characterize reservoir operational characteristics, achieve coordinated adaptation between power generation and ecological objectives, and make the scheduling scheme more closely aligned with the actual operating conditions of the reservoir.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling technology for water conservancy projects, and in particular to a method and system for multi-objective scheduling of smart reservoirs based on artificial intelligence. Background Technology

[0002] Conventional reservoir scheduling often relies on numerical fitting of single hydrological operation data, uses a single optimization model to carry out single-objective scheduling calculations for power generation or ecology, corrects scheduling parameters through manual experience, and only performs simple statistical analysis on basic data such as runoff and water level. It does not carry out spatiotemporal feature extraction processing on reservoir operation data, nor does it form a systematic feature vector that covers the periodic component of runoff, water level fluctuation mode, correlation coefficient of outflow, power generation load characteristics, and ecological water demand satisfaction index.

[0003] Existing scheduling methods can only rely on a single data dimension to carry out scheduling calculations, which cannot fully reflect the correlation characteristics of multi-dimensional operating parameters of the reservoir system. The scheduling model does not set up a step-by-step cascaded processing structure for runoff prediction, power generation optimization, and ecological scheduling. The scheduling objectives of power generation benefits and ecological water demand are prone to mismatch. The initial scheduling scheme is difficult to take into account multi-dimensional operating characteristics. The scheduling results do not match the actual operating conditions of the reservoir well. The ecological water demand satisfaction status and power generation operation requirements cannot be matched synchronously.

[0004] Multi-objective scheduling of reservoirs requires the extraction of spatiotemporal features from multi-dimensional historical operational data and the construction of a dedicated system feature vector. It also requires the construction of a deep neural network model consisting of a runoff prediction sub-network, a power generation optimization sub-network, and an ecological scheduling sub-network. The complete scheduling process of runoff prediction, power generation scheme generation, and ecological coordination optimization is completed according to a step-by-step processing logic, thereby meeting the actual needs of multi-objective collaborative scheduling of smart reservoirs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a smart reservoir multi-objective scheduling method and system based on artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a smart reservoir multi-objective scheduling method based on artificial intelligence, comprising:

[0007] A multi-dimensional historical operation data set of the target reservoir is collected, and spatiotemporal feature extraction processing is performed on the multi-dimensional historical operation data set to generate a reservoir system feature vector. The reservoir system feature vector includes runoff periodic components, water level fluctuation modes, outflow correlation coefficient, power generation load characteristics, and ecological water demand satisfaction index.

[0008] A multi-objective optimization model based on deep neural networks is constructed, which includes a runoff prediction sub-network, a power generation optimization sub-network, and an ecological scheduling sub-network.

[0009] The feature vector of the reservoir system is input into the multi-objective optimization model based on deep neural network, and the periodic components of the runoff are processed by the runoff prediction sub-network to generate the predicted runoff value for future periods.

[0010] The predicted inflow runoff is input into the power generation optimization subnetwork and processed in conjunction with the power generation load characteristics to generate an initial scheduling scheme that meets the power generation benefit target.

[0011] The initial scheduling scheme is input into the ecological scheduling sub-network and processed in conjunction with the ecological water demand satisfaction index to generate a comprehensive scheduling scheme that coordinates power generation and ecological objectives.

[0012] As a further aspect of the present invention, spatiotemporal feature extraction processing is performed on the multi-dimensional historical operational data set to generate a reservoir system feature vector, including:

[0013] The multi-dimensional historical operational data set includes inflow runoff sequence, water level change sequence, outflow sequence, power generation output sequence, and ecological water demand sequence;

[0014] The inflow runoff sequence is subjected to wavelet decomposition to separate runoff subsequences at multiple time scales, and the periodic component of runoff representing periodicity is extracted from the multiple time scale runoff subsequences.

[0015] The water level change sequence is subjected to empirical mode decomposition to obtain multiple intrinsic mode function components, and the water level fluctuation mode reflecting the main pattern of reservoir water level change is identified from the multiple intrinsic mode function components.

[0016] Cross-correlation analysis is performed on the outflow sequence and the inflow runoff sequence to calculate the outflow correlation coefficient, which reflects the degree of interdependence between the outflow sequence and the inflow runoff sequence.

[0017] Cluster analysis is performed on the power generation output sequence to identify different power generation mode categories, and the power generation load characteristics are constructed based on the statistical characteristics of each power generation mode category;

[0018] A matching degree analysis was performed on the ecological water demand sequence and the measured outflow sequence to calculate the ecological water demand satisfaction index, which reflects the degree of ecological demand satisfaction.

[0019] As a further aspect of the present invention, the construction of a multi-objective optimization model based on a deep neural network includes:

[0020] The runoff prediction subnetwork is constructed, which includes a long short-term memory layer and an attention mechanism layer, to capture the temporal dependencies and key temporal features in the periodic components of runoff;

[0021] The power generation optimization subnetwork is constructed, which includes a feedforward neural network layer and an objective function constraint layer, and is used to maximize power generation benefits under the constraint of the inflow runoff prediction value.

[0022] The ecological scheduling subnetwork is constructed, which includes a convolutional neural network layer and an adaptive weighted fusion layer, for evaluating and optimizing the ecological impact of the initial scheduling scheme;

[0023] The runoff prediction subnetwork, the power generation optimization subnetwork, and the ecological scheduling subnetwork are connected to form an end-to-end model architecture, and a multi-task learning method is used to jointly train the end-to-end model architecture.

[0024] As a further aspect of the present invention, the feature vector of the reservoir system is input into the multi-objective optimization model based on a deep neural network, and the periodic components of the runoff are processed by the runoff prediction sub-network to generate predicted values ​​of inflow runoff for future periods, including:

[0025] The reservoir system feature vector containing the periodic components of the runoff is input into the long short-term memory layer of the runoff prediction subnetwork.

[0026] In the long short-term memory layer, the periodic components of the runoff are time-series modeled to extract the trend characteristics and memory information of historical runoff;

[0027] The results output from the Long Short-Term Memory layer are input into the attention mechanism layer to calculate the importance weight of runoff information at different historical moments for future prediction.

[0028] Based on the aforementioned importance weights, runoff information from different historical moments is weighted and fused to generate preliminary runoff prediction results;

[0029] The preliminary runoff prediction results are post-processed and corrected to output the predicted runoff values ​​for future periods.

[0030] As a further aspect of the present invention, the predicted inflow runoff is input into the power generation optimization sub-network and processed in conjunction with the power generation load characteristics to generate an initial scheduling scheme that meets the power generation benefit objective, including:

[0031] The predicted inflow value for the future time period and the power generation load characteristics are jointly input into the feedforward neural network layer of the power generation optimization subnetwork;

[0032] In the feedforward neural network layer, the predicted inflow runoff for the future time period and the power generation load characteristics are subjected to nonlinear transformation and feature fusion to generate a high-dimensional feature representation;

[0033] The high-dimensional feature representation is input into the objective function constraint layer of the power generation optimization sub-network;

[0034] In the objective function constraint layer, the optimization objective is to maximize the power generation benefits obtained from the previous training, and the power generation scheduling optimization model is constructed with the reservoir water balance equation, reservoir capacity-water level relationship and outflow limit as constraints.

[0035] Solve the power generation dispatch optimization model to obtain the time period outflow plan and water level control curve that satisfy the constraints and tend to maximize power generation benefits, thus forming the initial dispatch scheme.

[0036] As a further aspect of the present invention, the initial scheduling scheme is input into the ecological scheduling sub-network and processed in conjunction with the ecological water demand satisfaction index to generate a comprehensive scheduling scheme that coordinates power generation and ecological objectives, including:

[0037] The time-period outflow plan and water level control curve in the initial scheduling scheme, as well as the ecological water demand satisfaction index, are input into the convolutional neural network layer of the ecological scheduling sub-network.

[0038] In the convolutional neural network layer, spatial and temporal features are extracted from the outflow plan for the specified time period and the ecological water demand satisfaction index to generate an ecological impact feature map.

[0039] The ecological impact feature map is input into the adaptive weighted fusion layer of the ecological scheduling sub-network;

[0040] In the adaptive weighted fusion layer, based on the predefined power generation target weights and ecological target weights, a multi-objective trade-off analysis is performed on the scheduling scheme reflected in the ecological impact feature map to generate coordinated scheduling scheme parameters.

[0041] Based on the coordinated scheduling scheme parameters, the initial scheduling scheme is adjusted to output the final comprehensive scheduling scheme that takes into account both power generation and ecological needs.

[0042] As a further aspect of the present invention, the method further includes the step of performing rolling optimization and feedback correction on the integrated scheduling scheme:

[0043] The comprehensive scheduling scheme is executed, and the actual operating data of the reservoir during the new scheduling cycle is collected in real time. The actual operating data includes actual inflow, actual outflow, actual power generation output, and actual downstream ecological indicators.

[0044] The actual operating data is compared with the predicted or planned values ​​corresponding to the integrated scheduling scheme to calculate the operating deviation data.

[0045] The operational deviation data is fed back to the multi-objective optimization model based on a deep neural network.

[0046] The parameters of the multi-objective optimization model based on deep neural networks are fine-tuned online using the aforementioned operational deviation data.

[0047] Using the fine-tuned multi-objective optimization model based on deep neural networks, an updated comprehensive scheduling scheme for the next scheduling cycle is generated based on the latest reservoir system status.

[0048] As a further aspect of the present invention, the method further includes the step of training and validating the multi-objective optimization model based on a deep neural network before scheduling decisions:

[0049] Prepare a training dataset containing multiple historical periods and multiple reservoir cases. The training dataset includes input features and corresponding ideal scheduling scheme labels.

[0050] The training dataset is divided into a model training subset and a model validation subset;

[0051] The model training subset is used to perform supervised training on the multi-objective optimization model based on deep neural networks. The internal parameters of the model are adjusted by the backpropagation algorithm so that the scheduling scheme output by the model approaches the ideal scheduling scheme label.

[0052] The model validation subset was used to evaluate the performance of the trained model, including the accuracy of runoff prediction, the degree of achievement of power generation benefits, and the degree of satisfaction of ecological needs.

[0053] Based on the evaluation results, adjust the model structure or hyperparameters, and repeat the training and validation steps until the model performance meets the preset deployment and application standards.

[0054] As a further aspect of the present invention, the power generation dispatch optimization model is solved to obtain a time-period outflow plan and water level control curve that satisfy the constraints and tend to maximize power generation benefits, thus forming the initial dispatch scheme, including:

[0055] The objective function and constraints in the power generation dispatch optimization model are transformed into a standard mathematical optimization problem form.

[0056] The standard mathematical optimization problem is solved iteratively using a sequential quadratic programming algorithm. In each iteration, a quadratic approximation of the objective function and a linear approximation of the constraints are constructed, and the resulting quadratic programming subproblem is solved.

[0057] Based on the search direction and step size obtained from solving the quadratic programming subproblem, update the current time period outflow plan and water level control curve variable values;

[0058] Determine whether the updated variable values ​​simultaneously satisfy the reservoir water balance equation, reservoir capacity-water level relationship, outflow limit constraints, and algorithm convergence criteria.

[0059] If satisfied, the current variable value is output as the optimal solution, forming the initial scheduling scheme that includes the outflow value and water level control elevation for a specific time period.

[0060] As a further aspect of the present invention, the present invention also includes an artificial intelligence-based intelligent reservoir multi-objective scheduling system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the artificial intelligence-based intelligent reservoir multi-objective scheduling method described above.

[0061] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0062] Spatiotemporal feature extraction and processing are performed on multi-dimensional historical operational data of the target reservoir to generate a reservoir system feature vector that includes periodic components of runoff, water level fluctuation modes, correlation coefficients of outflow, power generation load characteristics, and ecological water demand satisfaction indicators. This comprehensively quantifies the spatiotemporal variation characteristics of key parameters in reservoir operation, accurately reflecting the temporal variation characteristics of runoff, dynamic changes in water level, correlation characteristics of outflow parameters, operational characteristics of power generation load, and the status of ecological water security. Various feature parameters can fully reflect the operational status of the reservoir system, and the composition of the feature vector can directly match the input requirements of deep neural network models. The feature representation of operational data is more closely aligned with the inherent correlations of actual reservoir operation, the extraction and integration of parameter information is more systematic, and the feature expression of operational data is more targeted and complete.

[0063] A deep neural network multi-objective optimization model is constructed, consisting of a runoff prediction subnetwork, a power generation optimization subnetwork, and an ecological scheduling subnetwork. The subnetworks are cascaded to perform step-by-step calculations. The runoff prediction subnetwork predicts and extrapolates future inflows based on the periodic components of runoff. The power generation optimization subnetwork generates an initial scheduling scheme adapted to power generation targets by combining power generation load characteristics. The ecological scheduling subnetwork coordinates and optimizes the initial scheme by incorporating ecological water demand satisfaction indicators. The computational stages of each subnetwork are closely integrated, achieving a coherent computational flow from runoff prediction to power generation scheduling and then to ecological adaptation. Power generation and ecological scheduling targets are adapted and adjusted within a unified model system, reducing parameter conflicts between scheduling targets. The scheduling computation logic aligns with the actual execution process of multi-objective reservoir scheduling. The generated comprehensive scheduling scheme can simultaneously meet the management needs of power generation operation and ecological water use, improving the fit between the scheduling scheme and the actual operating conditions of the reservoir. Attached Figure Description

[0064] Figure 1 This is a flowchart of the intelligent reservoir multi-objective scheduling method based on artificial intelligence as described in this invention;

[0065] Figure 2 A flowchart for spatiotemporal feature extraction processing;

[0066] Figure 3 A flowchart for constructing a multi-objective optimization model for a deep neural network. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0068] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0069] See Figure 1 This invention provides a smart reservoir multi-objective scheduling method based on artificial intelligence. The overall implementation scheme of this method is as follows:

[0070] A multi-dimensional historical operational data set of the target reservoir is collected, containing various key time-series data accumulated during the reservoir's long-term operation. Spatiotemporal feature extraction processing is performed on this multi-dimensional historical operational data set. This processing aims to mine physically meaningful feature patterns from the raw data, generating a comprehensive reservoir system feature vector. This feature vector specifically includes the runoff periodic component, water level fluctuation mode, outflow correlation coefficient, power generation load characteristics, and ecological water demand satisfaction index, all extracted from the data. A multi-objective optimization model based on a deep neural network is constructed. This model is designed as a composite structure containing three specialized sub-networks: a runoff prediction sub-network, a power generation optimization sub-network, and an ecological scheduling sub-network. The reservoir system feature vector is input into the multi-objective optimization model based on the deep neural network. First, the runoff periodic component in the feature vector is processed by the runoff prediction sub-network to generate a predicted inflow value for a specific future period. This predicted inflow value is then input into the power generation optimization sub-network and combined with the power generation load characteristics in the feature vector for joint processing to generate an initial scheduling scheme that prioritizes maximizing power generation benefits. Subsequently, the initial scheduling scheme is input into the ecological scheduling sub-network and processed in conjunction with the ecological water demand satisfaction index in the feature vector. The initial scheme is evaluated and adjusted in an ecological dimension, and finally a comprehensive scheduling scheme that can coordinate power generation benefits and ecological protection goals is generated.

[0071] In one embodiment of the present invention, the multi-dimensional historical operational data set specifically includes inflow runoff sequence, water level change sequence, outflow sequence, power generation output sequence, and ecological water demand sequence. See also... Figure 2The inflow sequence is processed by wavelet decomposition to separate the original runoff sequence into multiple runoff subsequences at different time scales. Periodicity components characterizing the reservoir's inflow cycle are extracted from these decomposed runoff subsequences. The water level change sequence is processed by empirical mode decomposition to obtain a series of intrinsic mode function (IMF) components from high to low frequencies. Water level fluctuation modes reflecting the main patterns of reservoir water level changes are identified from these IMF components. Cross-correlation analysis is performed on the outflow sequence and the inflow runoff sequence. By calculating the correlation coefficients of the two sequences at different time delays, the outflow correlation coefficient, reflecting the degree of interdependence between the two sequences, is obtained. Cluster analysis is performed on the power generation output sequence to identify different typical power generation mode categories. Power generation load characteristics are constructed based on the statistical characteristics of each power generation mode category. A matching degree analysis was performed on the ecological water demand sequence and the measured outflow sequence. By calculating the ratio of the measured outflow to the ecological water demand sequence within a specific time period, an ecological water demand satisfaction index reflecting the degree of ecological demand satisfaction was obtained. The runoff periodic component, water level fluctuation mode, outflow correlation coefficient, power generation load characteristics, and ecological water demand satisfaction index obtained from the above processing together constitute the reservoir system feature vector.

[0072] In practical implementation, spatiotemporal feature extraction processing is carried out on a multi-dimensional historical operational data set. Taking the ten-year historical operational data of a reservoir as an example scenario, the multi-dimensional historical operational data set specifically includes inflow runoff sequence, water level change sequence, outflow sequence, power generation output sequence, and ecological water demand sequence. The inflow runoff sequence is a daily-scale observation sequence, the water level change sequence is an hourly-scale observation sequence, the outflow sequence is a daily-scale control value sequence, the power generation output sequence is an hourly-scale recorded value sequence, and the ecological water demand sequence is a daily-scale demand threshold sequence. Wavelet decomposition processing is performed on the inflow runoff sequence. A five-level discrete wavelet transform is applied to the inflow runoff sequence using the Daubechies wavelet basis to separate multiple time-scale runoff subsequences, including high-frequency detail components D1 and D2 and low-frequency approximation component A5. From the runoff subsequences of multiple time scales, the runoff periodic component characterizing the periodicity is extracted. The main period length of the runoff periodic component is determined by calculating the peak position of the autocorrelation coefficient of the low-frequency approximation component A5 between different years, and the period intensity is quantified by wavelet variance analysis. Empirical mode decomposition (EMD) was performed on the water level change sequence. Through a screening process, the water level change sequence was decomposed into eight intrinsic mode function (IMF) components and one residual term. From the multiple IMF components, the water level fluctuation mode reflecting the main pattern of reservoir water level change was identified. The water level fluctuation mode was reconstructed by superimposing the top three IMF components with an energy ratio of more than 60%. Its fluctuation characteristics were described by the number of extreme points and the number of zero crossings.

[0073] In some embodiments, cross-correlation analysis is performed on the outflow and inflow runoff sequences to calculate the outflow correlation coefficient, which reflects the degree of interdependence between the two sequences. Specifically, a maximum time lag of 30 days is set, and the cross-correlation coefficient between the inflow and outflow sequences is calculated daily at time lag k. The outflow correlation coefficient is taken as the maximum absolute value of the cross-correlation coefficients across all time lags, and its calculation formula is as follows:

[0074] in: The Pearson correlation coefficient is represented by a time lag of k days, where k ranges from -30 to 30. Cluster analysis is performed on the power generation output sequence using the K-means clustering algorithm, with Euclidean distance as the metric. The daily load curves of the power generation output sequence are clustered into three output pattern categories. Based on the statistical characteristics of each output pattern category, a power generation load feature is constructed. This feature is a vector containing the cluster center curve, the average output value within the cluster, and the frequency of occurrence of that category for each output pattern category. Optionally, the number of clusters can be determined using the elbow rule.

[0075] In practical implementation, a matching degree analysis is performed on the ecological water demand sequence and the measured outflow sequence to calculate the ecological water demand satisfaction index, which reflects the degree of ecological demand satisfaction. Within a scheduling year, using ten-day periods as the basic calculation unit, the measured outflow sequence is compared ten-day periods with the corresponding ecological base flow demand values ​​in the ecological water demand sequence. The ecological water demand satisfaction index is defined as the proportion of days when the measured outflow is not lower than the ecological base flow demand to the total number of days in that ten-day period. The satisfaction ratios for all 36 ten-day periods constitute a satisfaction sequence, and the annual average of this sequence is the annual ecological water demand satisfaction index. Optionally, the ecological water demand sequence can further include pulse flow demand during the fish breeding season. In this case, the matching degree analysis needs to add a conformity assessment of the timing, duration, and flow magnitude of the pulse event. In some embodiments, the final generated reservoir system feature vector is a structured data object that sequentially contains an array of runoff periodic components of length L, an IMF component parameter set describing the water level fluctuation mode, a scalar value of the outflow correlation coefficient, a clustering parameter vector characterizing the power generation load characteristics, and the ecological water demand satisfaction index value.

[0076] In one embodiment of the present invention, see [reference] Figure 3The following subnetworks are constructed: a runoff prediction subnetwork, comprising a long short-term memory (LSTM) layer and an attention mechanism layer, designed to capture the long and short-term temporal dependencies and key temporal features in the periodic components of runoff; a power generation optimization subnetwork, comprising a feedforward neural network layer and an objective function constraint layer, designed to solve an optimization problem maximizing power generation benefits under the constraints of the predicted inflow runoff; and an ecological scheduling subnetwork, comprising a convolutional neural network layer and an adaptive weighted fusion layer, designed to evaluate and optimize the impact of scheduling schemes on the downstream ecological environment. The runoff prediction subnetwork, the power generation optimization subnetwork, and the ecological scheduling subnetwork are connected to form an end-to-end model architecture, which is then jointly trained using a multi-task learning method. When generating predicted values, the reservoir system feature vector containing the periodic components of runoff is input into the LSM layer of the runoff prediction subnetwork. In the LSM layer, time-series modeling of the periodic components of runoff is performed, extracting trend features and memory information contained in historical runoff data. The output of the Long Short-Term Memory (LSTM) layer is input into the attention mechanism layer to calculate the importance weight of runoff information at different historical times for future prediction times. Based on the calculated importance weights, runoff information from different historical times is weighted and fused to generate a preliminary runoff prediction result. The preliminary runoff prediction result is post-processed and corrected to output the final predicted runoff value for the future period.

[0077] In practical implementation, constructing a multi-objective optimization model based on deep neural networks involves the structural definition and connection of three specialized sub-networks. The runoff prediction sub-network includes a bidirectional long short-term memory (LSTM) layer with 128 hidden units and an attention mechanism layer with 64 key-value pairs. The LSM layer captures the forward and backward temporal dependencies in the periodic components of runoff, while the attention mechanism layer identifies and weights the criticality of different historical time steps in predicting future inflow runoff. The power generation optimization sub-network includes a three-layer feedforward neural network and an objective function constraint layer integrating linear and nonlinear constraint expressions. The three layers of the feedforward neural network have 256, 128, and 64 neurons, respectively. The objective function constraint layer expresses the goal of maximizing power generation benefits as a function of power generation and reservoir head. The ecological scheduling sub-network includes a convolutional neural network layer with one-dimensional convolutional kernels and an adaptive weighted fusion layer with learnable weight parameters. The convolutional neural network layer uses three temporal convolutional kernels of different widths to extract multi-scale temporal features. The adaptive weighted fusion layer receives the ecological impact feature map from the convolutional neural network layer and performs multi-objective weighting. An end-to-end model architecture is formed by connecting the runoff prediction subnetwork, the power generation optimization subnetwork, and the ecological scheduling subnetwork. The connection method is to use the output of the runoff prediction subnetwork as one of the inputs of the power generation optimization subnetwork, and to use the output of the power generation optimization subnetwork and the ecological water demand satisfaction index as inputs of the ecological scheduling subnetwork. A multi-task learning method is used to jointly train the end-to-end model architecture. The loss function of multi-task learning is a weighted sum of the mean square error of runoff prediction, the negative value of power generation benefit, and the ecological deviation penalty term.

[0078] In some embodiments, a reservoir system feature vector containing runoff periodic components is input into the long short-term memory (LSTM) layer of the runoff prediction subnetwork. The runoff periodic components are organized as continuous time-series data, for example, using data from the past 360 days as the input sequence. In the LSTM layer, time-series modeling of the runoff periodic components is performed, extracting trend features and memory information from historical runoff. The bidirectional LSTM layer processes the input sequence from both forward and reverse directions, concatenating the final hidden states from both directions. The output of the LSTM layer is input into an attention mechanism layer to calculate the importance weights of runoff information at different historical moments for future prediction. These importance weights are calculated using a single-layer perceptron, with the input being the hidden states of the LSTM layer at the corresponding time step. Based on these importance weights, runoff information from different historical moments is weighted and fused to generate a preliminary runoff prediction result. The weighted fusion is achieved by calculating the weighted sum of the hidden states at all historical moments, expressed by the formula:

[0079]

[0080] in: It is the generated context vector. It is the first The attention weight of each historical moment Is the Long Short-Term Memory layer in the first... The hidden state at a given moment This represents the total number of time steps in the input sequence. Optionally, the attention weights can be normalized using the softmax function.

[0081] In practice, the preliminary runoff forecast results are post-processed and corrected to output the predicted inflow runoff values ​​for future periods. This post-processing correction is implemented using a fully connected layer, which integrates the context vector... This is mapped to a sequence of predicted inflow values ​​for the next 7 or 30 days. It's understood that the number of units in the Long Short-Term Memory (LSTM) layer, the dimension of the attention mechanism layer, and the length of the predicted future period are adjustable hyperparameters that need to be set according to the specific reservoir's data characteristics and scheduling cycle. In some embodiments, the runoff prediction subnetwork can be pre-trained separately, using historical runoff data to train the network parameters in a supervised learning manner, with the training objective being to minimize the root mean square error between the predicted and actual values. Optionally, the parameters of the runoff prediction subnetwork can also be synchronously updated using a multi-task loss function during end-to-end joint training. It's understood that the weight distribution generated by the attention mechanism layer can be used to explain model decisions, and the historical moments corresponding to high weights can be interpreted as key historical periods influencing future runoff.

[0082] In one embodiment of the present invention, this embodiment details how to generate an initial scheduling scheme through a power generation optimization sub-network and how to solve the power generation scheduling optimization model. The predicted inflow values ​​for future periods and the power generation load characteristics are input into the feedforward neural network layer of the power generation optimization sub-network. In the feedforward neural network layer, the predicted inflow values ​​for future periods and the power generation load characteristics undergo nonlinear transformation and feature fusion to generate a high-dimensional feature representation for subsequent optimization calculations. This high-dimensional feature representation is input into the objective function constraint layer of the power generation optimization sub-network. In the objective function constraint layer, maximizing the power generation benefit obtained from previous training is used as the optimization objective, and the reservoir water balance equation, reservoir capacity-water level relationship, and outflow limit are used as constraints to construct a power generation scheduling optimization model. Solving the power generation scheduling optimization model yields the outflow plan and water level control curve for the period that satisfy the constraints and tends to maximize power generation benefit, forming the initial scheduling scheme. The specific solution process includes: transforming the objective function and constraints in the power generation scheduling optimization model into a standard mathematical optimization problem form. A sequential quadratic programming algorithm is used to iteratively solve the standard mathematical optimization problem. In each iteration, a quadratic approximation of the objective function and a linear approximation of the constraints are constructed, and the resulting quadratic programming subproblem is solved. Based on the search direction and step size obtained from solving the quadratic programming subproblem, the current outflow plan and water level control curve variable values ​​are updated. It is determined whether the updated variable values ​​simultaneously satisfy the reservoir water balance equation, the reservoir capacity-water level relationship, the outflow limit constraints, and the algorithm convergence criteria. If satisfied, the current variable value is output as the optimal solution, forming the initial scheduling scheme containing the specific outflow values ​​and water level control elevations for each time period.

[0083] In practical implementation, the predicted inflow runoff for future periods and the power generation load characteristics are jointly input into the feedforward neural network layer of the power generation optimization sub-network. The predicted inflow runoff for future periods is a sequence of length N, representing the predicted inflow volume for the next N scheduling periods. The power generation load characteristics are a vector containing cluster centers, average output, and frequency information. The first layer of the feedforward neural network receives the concatenated vector of these two inputs. In the feedforward neural network layer, the predicted inflow runoff for future periods and the power generation load characteristics undergo nonlinear transformation and feature fusion to generate a high-dimensional feature representation. The nonlinear transformation is implemented using the ReLU activation function, and feature fusion is completed in the deep layers of the network through fully connected operations. The final high-dimensional feature representation is a vector of dimension M. This high-dimensional feature representation is input into the objective function constraint layer of the power generation optimization sub-network, serving as the initial point or part of the parameterized expression of the optimization problem. In the objective function constraint layer, maximizing the power generation benefit obtained from previous training is the optimization objective. The power generation benefit is typically expressed as maximizing the total power generation during the scheduling period. A power generation scheduling optimization model is constructed using the reservoir water balance equation, reservoir capacity-water level relationship, and outflow limit as constraints. Solve the power generation dispatch optimization model to obtain the time period outflow plan and water level control curve that satisfy the constraints and tend to maximize power generation benefits, and form the initial dispatch scheme. The initial dispatch scheme includes the planned outflow value of each time period in the next N time periods and the reservoir water level control value at the end of the time period.

[0084] In some embodiments, the specific steps for solving the power generation dispatch optimization model include transforming the objective function and constraints in the power generation dispatch optimization model into a standard mathematical optimization problem form, which is to minimize a scalar function with equality and inequality constraints. A sequential quadratic programming algorithm is used to iteratively solve the standard mathematical optimization problem. In each iteration, the sequential quadratic programming algorithm constructs a quadratic approximation of the objective function and a linear approximation of the constraints, and solves the resulting quadratic programming subproblem. The solution to the quadratic programming subproblem serves as the search direction for the current iteration point. Based on the search direction and step size obtained from solving the quadratic programming subproblem, the current time-period outflow plan and water level control curve variable values ​​are updated. The step size is determined through a one-dimensional linear search to ensure a decrease in the objective function. It is then determined whether the updated variable values ​​simultaneously satisfy the reservoir water balance equation, the reservoir capacity-water level relationship, the outflow restriction constraints, and the algorithm convergence criteria. The convergence criteria include the variable change being less than a threshold or the objective function improvement being less than a threshold. If satisfied, the current variable value is output as the optimal solution, forming an initial scheduling scheme that includes the outflow rate and water level control elevation for specific time periods. The specific values ​​can be presented in tabular form. See Table 1, which shows a data segment of an initial scheduling scheme containing 7 time periods:

[0085] Table 1: Initial Scheduling Scheme

[0086] 1 150.2 175.0 2 162.8 174.7 3 158.3 174.5 4 170.5 174.1 5 165.0 173.9 6 155.7 173.8 7 160.1 173.6

[0087] In practical implementation, the objective function of the power generation dispatch optimization model can be expressed as maximizing the total power generation during the dispatch period, and its mathematical expression is:

[0088]

[0089] in: Indicates the total number of time periods during the scheduling period. It is the overall output coefficient. It is the first Power generation flow during the period It is the first Average hydropower head over the period This refers to the length of the time period. The constraints of the reservoir water balance equation are expressed as follows: , It is the first Initial storage capacity at the beginning of the period It is the first Inflow rate for a given period (based on predicted inflow runoff for future periods). It is the first The outflow rate during a given period. This can be understood in relation to the reservoir's capacity and water level. Determined by the reservoir characteristic curve, it is usually given in the form of a lookup table or a fitted function. The outflow limit constraint is expressed as follows: ,in and These represent the minimum and maximum allowable outflow rates determined by technical conditions. Optionally, the objective function constraint layer should also consider other engineering constraints such as power plant output limits and water level fluctuation limits when constructing the power generation dispatch optimization model. In some embodiments, the sequential quadratic programming algorithm can use the effective set method or the interior point method to solve the quadratic programming subproblem.

[0090] In one embodiment of the present invention, the time-period outflow plan and water level control curve from the initial scheduling scheme, along with the ecological water demand satisfaction index, are input into the convolutional neural network layer of the ecological scheduling sub-network. In the convolutional neural network layer, spatial and temporal features are extracted from the time-period outflow plan and the ecological water demand satisfaction index to generate an ecological impact feature map characterizing the ecological impact of the scheduling scheme. The ecological impact feature map is input into the adaptive weighted fusion layer of the ecological scheduling sub-network. In the adaptive weighted fusion layer, based on predefined power generation target weights and ecological target weights, a multi-objective trade-off analysis is performed on the scheduling scheme reflected in the ecological impact feature map to generate coordinated scheduling scheme parameters. Based on the coordinated scheduling scheme parameters, the initial scheduling scheme is adjusted, and the final comprehensive scheduling scheme, which considers both power generation and ecological needs, is output. The comprehensive scheduling scheme is executed, and real-time operational data of the reservoir during the new scheduling cycle is collected. This operational data includes actual inflow, actual outflow, actual power generation output, and actual downstream ecological indicators. The actual operational data is compared with the predicted or planned values ​​corresponding to the integrated scheduling scheme to calculate the operational deviation data. This operational deviation data is then fed back to the deep neural network-based multi-objective optimization model. The parameters of the deep neural network-based multi-objective optimization model are fine-tuned online using the operational deviation data. Using the fine-tuned deep neural network-based multi-objective optimization model, and based on the latest reservoir system state, an updated integrated scheduling scheme for the next scheduling cycle is generated.

[0091] In practical implementation, the time-period outflow plan and water level control curve from the initial scheduling scheme, along with the ecological water demand satisfaction index, are input into the convolutional neural network layer of the ecological scheduling sub-network. The initial scheduling scheme includes outflow and water level sequences for the next N time periods, and the ecological water demand satisfaction index contains vectors representing the degree of ecological demand satisfaction in historical time periods. The convolutional neural network layer is designed with multiple one-dimensional convolutional kernels to process the input sequences in parallel. Within the convolutional neural network layer, spatial and temporal features are extracted from the time-period outflow plan and the ecological water demand satisfaction index to generate an ecological impact feature map. Specifically, the outflow plan sequence and the ecological water demand satisfaction index sequence are concatenated along the channel dimension to form a multi-channel input tensor. Multiple one-dimensional convolutional kernels slide along the time dimension on this tensor to perform convolution operations and extract local and global dependency patterns. The output is the ecological impact feature map. This ecological impact feature map is then input into the adaptive weighted fusion layer of the ecological scheduling sub-network. The ecological impact feature map reflects the potential pressure distribution on the ecosystem across different scheduling schemes in the temporal dimension. In the adaptive weighted fusion layer, a multi-objective trade-off analysis is performed on the scheduling scheme reflected in the ecological impact feature map based on predefined power generation target weights and ecological target weights. This generates coordinated scheduling scheme parameters. The power generation target weights and ecological target weights are set by decision-makers or learned from historical data. The multi-objective trade-off analysis is implemented through a fully connected network that takes the ecological impact feature map as input and outputs the adjustment amount of the scheduling scheme. Based on the coordinated scheduling scheme parameters, the initial scheduling scheme is adjusted, outputting the final comprehensive scheduling scheme that considers both power generation and ecological needs. The adjustment typically involves smoothing the outflow process or reducing peak flows while meeting ecological constraints.

[0092] In some embodiments, the multi-objective trade-offs performed by the adaptive weighted fusion layer can be quantified as solving an optimization problem whose objective is to minimize the combined loss between power generation and ecological objectives, expressed as:

[0093]

[0094] in: It is the weight of the power generation target. It is the weight of ecological goals. It is a loss item that is negatively correlated with power generation efficiency. It is the ecological deviation loss item. and All are calculated from the ecological impact feature map using different mapping functions. Optional, power generation target weights. Weight of ecological goals It can be set to change dynamically, for example, based on the current water level of the reservoir or the season. In a simplified example scenario covering seven time periods, a comparison of some key data between the initial scheduling scheme and the integrated scheduling scheme adjusted by the ecological scheduling sub-network is shown in Table 2:

[0095] Table 2: Comparison of Initial Scheduling Scheme and Integrated Scheduling Scheme

[0096] 1 150.2 155.5 175.0 174.9 2 162.8 160.1 174.7 174.6 3 158.3 161.0 174.5 174.5 4 170.5 168.2 174.1 174.2 5 165.0 163.5 173.9 174.0 6 155.7 158.0 173.8 173.9 7 160.1 159.8 173.6 173.7

[0097] In practical implementation, the steps for rolling optimization and feedback correction of the integrated scheduling scheme include executing the integrated scheduling scheme and collecting real-time actual operating data of the reservoir in the new scheduling cycle. The actual operating data includes actual inflow, actual outflow, actual power generation output, and actual downstream ecological indicators. The actual operating data is compared with the predicted or planned values ​​corresponding to the integrated scheduling scheme to calculate operating deviation data. For example, the relative error sequence between actual and predicted inflow is calculated, and the absolute error sequence between actual and planned outflow is calculated. The operating deviation data is fed back to a multi-objective optimization model based on a deep neural network. The operating deviation data is organized into a structure consistent with the model's input feature dimensions. The parameters of the multi-objective optimization model based on the deep neural network are fine-tuned online using the operating deviation data. Online fine-tuning employs a mini-batch gradient descent method, aiming to minimize the loss function constituted by the operating deviation, and making small updates to some or all of the model's parameters. Using the fine-tuned multi-objective optimization model based on the deep neural network, and based on the latest reservoir system state, an updated integrated scheduling scheme for the next scheduling cycle is generated, thus forming a closed-loop, adaptive scheduling decision-making process. It is understandable that the rolling optimization cycle can be consistent with the time period of the scheduling scheme, for example, generating a scheduling scheme for the next 7 days daily. In some embodiments, actual downstream ecological indicators can be obtained through hydrological and water quality sensors at downstream monitoring sections, such as dissolved oxygen concentration or fish activity monitoring data. Optionally, operational deviation data can be smoothed before feedback to eliminate the interference of random errors. It is understood that the frequency and learning rate of online fine-tuning need to be carefully set to avoid catastrophic forgetting of the model when adapting to new data.

[0098] In one embodiment of the present invention, this embodiment details the steps for training and validating a multi-objective optimization model based on a deep neural network before scheduling decisions are made. A training dataset containing multiple historical periods and multiple reservoir cases is prepared. The training dataset includes input features and corresponding ideal scheduling scheme labels. The training dataset is divided into a model training subset and a model validation subset. The model training subset is used to perform supervised training on the multi-objective optimization model based on the deep neural network. The internal parameters of the model are adjusted using a backpropagation algorithm so that the scheduling scheme output by the model approaches the ideal scheduling scheme label. The performance of the trained model is evaluated using the model validation subset. The evaluation includes the accuracy of runoff prediction, the degree of achievement of power generation benefits, and the degree of satisfaction of ecological needs. Based on the evaluation results, the model structure or hyperparameters are adjusted, and the training and validation steps are repeated until the model performance meets the preset deployment and application standards.

[0099] In practical implementation, the steps of training and validating the multi-objective optimization model based on deep neural networks before scheduling decisions include preparing a training dataset containing multiple historical periods and multiple reservoir cases. The training dataset covers at least ten years of historical hydrological years and includes operational data from no fewer than five reservoirs with different regulation capabilities. The training dataset includes input features and corresponding ideal scheduling scheme labels. The input features are the reservoir system feature vectors generated in the aforementioned embodiments, and the ideal scheduling scheme labels are the outflow and water level sequences corresponding to scheduling schemes that have been verified as superior through post-analysis in the same historical records. The training dataset is divided into a model training subset and a model validation subset, which are divided in chronological order or randomly. The data volume ratio of the model training subset to the model validation subset is usually set to 7:3 to ensure that the model fully learns historical patterns and has a basis for evaluating generalization ability. Supervised training of a deep neural network-based multi-objective optimization model is performed using a subset of the model training data. The model's internal parameters are adjusted using backpropagation. Backpropagation calculates the gradient of the loss function based on the difference between the model's output scheduling scheme and the ideal scheduling scheme label, and then backpropagates layer by layer to update the network weights and biases, making the model's output scheduling scheme approximate the ideal scheduling scheme label. In some embodiments, the loss function during training is designed as a multi-objective weighted form, expressed as:

[0100]

[0101] in: It is the total loss. It is the prediction mean square error loss of the runoff prediction subnetwork. These are losses related to power generation efficiency (e.g., negative power generation). It is an ecological deviation loss. , and These are hyperparameters used to balance the weights of the three loss terms. The process of adjusting the model's internal parameters using the backpropagation algorithm is to minimize the total loss. The process involves evaluating the performance of the trained model using a model validation subset. The evaluation includes runoff prediction accuracy, power generation benefit achievement, and ecological demand satisfaction. Runoff prediction accuracy is assessed by calculating the Nash efficiency coefficients of predicted and actual runoff on the validation set. Power generation benefit achievement is assessed by comparing the total power generation of the model's scheduling scheme and the validation set's labeled scheme. Ecological demand satisfaction is assessed by comparing the degree to which the two schemes guarantee downstream ecological baseflow. These evaluation results are used to determine whether the model is overfitting or underfitting, and the degree of coordination among the sub-objectives.

[0102] In practice, based on the evaluation results, the model structure or hyperparameters are adjusted, and the training and validation steps are repeated until the model performance meets the preset deployment and application standards. Adjusting the model structure may involve increasing or decreasing the number of layers in each sub-network, adjusting the number of neurons in the hidden layers, or modifying the number of attention mechanism heads. Adjusting hyperparameters may involve changing the learning rate, batch size, and loss function weights. , , Or other balancing parameters in multi-task learning. The repeated training and validation steps are an iterative process; after each adjustment, the same training subset of the model is retrained, and the performance is re-evaluated using the validation subset. Predefined deployment and application criteria are a predefined set of performance thresholds, such as requiring a Nash efficiency coefficient for runoff prediction higher than 0.8, a power generation benefit achievement rate of no less than 95% of the historical best, and an ecological demand satisfaction rate of no less than 90%. Optionally, the model validation subset can be further subdivided into a validation set and a test set. The validation set is used to adjust hyperparameters and select the model, while the test set is used to finally evaluate the model's generalization performance on unseen data. In some embodiments, the training process can use an early stopping strategy to prevent overfitting; when the model's performance on the validation subset no longer improves over several consecutive training cycles, training is stopped, and the best-performing model parameters are retained. It is understood that the model that ultimately meets the deployment and application criteria will be saved and used to generate integrated reservoir scheduling schemes online.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A smart reservoir multi-objective scheduling method based on artificial intelligence, characterized in that, The method includes: A multi-dimensional historical operation data set of the target reservoir is collected, and spatiotemporal feature extraction processing is performed on the multi-dimensional historical operation data set to generate a reservoir system feature vector. The reservoir system feature vector includes runoff periodic components, water level fluctuation modes, outflow correlation coefficient, power generation load characteristics, and ecological water demand satisfaction index. A multi-objective optimization model based on deep neural networks is constructed, which includes a runoff prediction sub-network, a power generation optimization sub-network, and an ecological scheduling sub-network. The feature vector of the reservoir system is input into the multi-objective optimization model based on deep neural network, and the periodic components of the runoff are processed by the runoff prediction sub-network to generate the predicted runoff value for future periods. The predicted inflow runoff is input into the power generation optimization subnetwork and processed in conjunction with the power generation load characteristics to generate an initial scheduling scheme that meets the power generation benefit target. The initial scheduling scheme is input into the ecological scheduling sub-network and processed in conjunction with the ecological water demand satisfaction index to generate a comprehensive scheduling scheme that coordinates power generation and ecological objectives.

2. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 1, characterized in that, The multi-dimensional historical operational data set is subjected to spatiotemporal feature extraction processing to generate a reservoir system feature vector, including: The multi-dimensional historical operational data set includes inflow runoff sequence, water level change sequence, outflow sequence, power generation output sequence, and ecological water demand sequence; The inflow runoff sequence is subjected to wavelet decomposition to separate runoff subsequences at multiple time scales, and the periodic component of runoff representing periodicity is extracted from the multiple time scale runoff subsequences. The water level change sequence is subjected to empirical mode decomposition to obtain multiple intrinsic mode function components, and the water level fluctuation mode reflecting the main pattern of reservoir water level change is identified from the multiple intrinsic mode function components. Cross-correlation analysis is performed on the outflow sequence and the inflow runoff sequence to calculate the outflow correlation coefficient, which reflects the degree of interdependence between the outflow sequence and the inflow runoff sequence. Cluster analysis is performed on the power generation output sequence to identify different power generation mode categories, and the power generation load characteristics are constructed based on the statistical characteristics of each power generation mode category; A matching degree analysis was performed on the ecological water demand sequence and the measured outflow sequence to calculate the ecological water demand satisfaction index, which reflects the degree of ecological demand satisfaction.

3. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 1, characterized in that, The construction of the multi-objective optimization model based on deep neural networks includes: The runoff prediction subnetwork is constructed, which includes a long short-term memory layer and an attention mechanism layer, to capture the temporal dependencies and key temporal features in the periodic components of runoff; The power generation optimization subnetwork is constructed, which includes a feedforward neural network layer and an objective function constraint layer, and is used to maximize power generation benefits under the constraint of the inflow runoff prediction value. The ecological scheduling subnetwork is constructed, which includes a convolutional neural network layer and an adaptive weighted fusion layer, for evaluating and optimizing the ecological impact of the initial scheduling scheme; The runoff prediction subnetwork, the power generation optimization subnetwork, and the ecological scheduling subnetwork are connected to form an end-to-end model architecture, and a multi-task learning method is used to jointly train the end-to-end model architecture.

4. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 3, characterized in that, The feature vector of the reservoir system is input into the multi-objective optimization model based on a deep neural network. The periodic components of the runoff are processed by the runoff prediction sub-network to generate predicted runoff values ​​for future periods, including: The reservoir system feature vector containing the periodic components of the runoff is input into the long short-term memory layer of the runoff prediction subnetwork. In the long short-term memory layer, the periodic components of the runoff are time-series modeled to extract the trend characteristics and memory information of historical runoff; The results output from the Long Short-Term Memory layer are input into the attention mechanism layer to calculate the importance weight of runoff information at different historical moments for future prediction. Based on the aforementioned importance weights, runoff information from different historical moments is weighted and fused to generate preliminary runoff prediction results; The preliminary runoff prediction results are post-processed and corrected to output the predicted runoff values ​​for future periods.

5. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 4, characterized in that, The predicted inflow runoff is input into the power generation optimization subnetwork and processed in conjunction with the power generation load characteristics to generate an initial scheduling scheme that meets the power generation benefit target, including: The predicted inflow value for the future time period and the power generation load characteristics are jointly input into the feedforward neural network layer of the power generation optimization subnetwork; In the feedforward neural network layer, the predicted inflow runoff for the future time period and the power generation load characteristics are subjected to nonlinear transformation and feature fusion to generate a high-dimensional feature representation; The high-dimensional feature representation is input into the objective function constraint layer of the power generation optimization sub-network; In the objective function constraint layer, the optimization objective is to maximize the power generation benefits obtained from the previous training, and the power generation scheduling optimization model is constructed with the reservoir water balance equation, reservoir capacity-water level relationship and outflow limit as constraints. Solve the power generation dispatch optimization model to obtain the time period outflow plan and water level control curve that satisfy the constraints and tend to maximize power generation benefits, thus forming the initial dispatch scheme.

6. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 5, characterized in that, The initial scheduling scheme is input into the ecological scheduling sub-network and processed in conjunction with the ecological water demand satisfaction index to generate a comprehensive scheduling scheme that coordinates power generation and ecological objectives, including: The time-period outflow plan and water level control curve in the initial scheduling scheme, as well as the ecological water demand satisfaction index, are input into the convolutional neural network layer of the ecological scheduling sub-network. In the convolutional neural network layer, spatial and temporal features are extracted from the outflow plan for the specified time period and the ecological water demand satisfaction index to generate an ecological impact feature map. The ecological impact feature map is input into the adaptive weighted fusion layer of the ecological scheduling sub-network; In the adaptive weighted fusion layer, based on the predefined power generation target weights and ecological target weights, a multi-objective trade-off analysis is performed on the scheduling scheme reflected in the ecological impact feature map to generate coordinated scheduling scheme parameters. Based on the coordinated scheduling scheme parameters, the initial scheduling scheme is adjusted to output the final comprehensive scheduling scheme that takes into account both power generation and ecological needs.

7. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 6, characterized in that, The method further includes the steps of rolling optimization and feedback correction of the integrated scheduling scheme: The comprehensive scheduling scheme is executed, and the actual operating data of the reservoir during the new scheduling cycle is collected in real time. The actual operating data includes actual inflow, actual outflow, actual power generation output, and actual downstream ecological indicators. The actual operating data is compared with the predicted or planned values ​​corresponding to the integrated scheduling scheme to calculate the operating deviation data. The operational deviation data is fed back to the multi-objective optimization model based on a deep neural network. The parameters of the multi-objective optimization model based on deep neural networks are fine-tuned online using the aforementioned operational deviation data. Using the fine-tuned multi-objective optimization model based on deep neural networks, an updated comprehensive scheduling scheme for the next scheduling cycle is generated based on the latest reservoir system status.

8. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 1, characterized in that, The method further includes the steps of training and validating the deep neural network-based multi-objective optimization model before scheduling decisions: Prepare a training dataset containing multiple historical periods and multiple reservoir cases. The training dataset includes input features and corresponding ideal scheduling scheme labels. The training dataset is divided into a model training subset and a model validation subset; The model training subset is used to perform supervised training on the multi-objective optimization model based on deep neural networks. The internal parameters of the model are adjusted by the backpropagation algorithm so that the scheduling scheme output by the model approaches the ideal scheduling scheme label. The model validation subset was used to evaluate the performance of the trained model, including the accuracy of runoff prediction, the degree of achievement of power generation benefits, and the degree of satisfaction of ecological needs. Based on the evaluation results, adjust the model structure or hyperparameters, and repeat the training and validation steps until the model performance meets the preset deployment and application standards.

9. The intelligent reservoir multi-objective scheduling method based on artificial intelligence according to claim 5, characterized in that, Solving the power generation dispatch optimization model yields the time-period outflow plan and water level control curve that satisfy the constraints and tend to maximize power generation benefits, forming the initial dispatch scheme, including: The objective function and constraints in the power generation dispatch optimization model are transformed into a standard mathematical optimization problem form. The standard mathematical optimization problem is solved iteratively using a sequential quadratic programming algorithm. In each iteration, a quadratic approximation of the objective function and a linear approximation of the constraints are constructed, and the resulting quadratic programming subproblem is solved. Based on the search direction and step size obtained from solving the quadratic programming subproblem, update the current time period outflow plan and water level control curve variable values; Determine whether the updated variable values ​​simultaneously satisfy the reservoir water balance equation, reservoir capacity-water level relationship, outflow limit constraints, and algorithm convergence criteria. If satisfied, the current variable value is output as the optimal solution, forming the initial scheduling scheme that includes the outflow value and water level control elevation for a specific time period.

10. A smart reservoir multi-objective scheduling system based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the AI-based smart reservoir multi-objective scheduling method as described in any one of claims 1 to 9.

Images