A reservoir operation state monitoring method and system based on deep learning

Abstract

This application provides a method and system for monitoring the operational status of reservoirs based on deep learning, relating to the field of water conservancy project operation monitoring. The method includes: collecting multi-source time-series data on reservoir operation, including hydrological time-series data, engineering time-series data, and environmental time-series data; constructing a multi-source feature sequence of the reservoir's operational status based on the multi-source time-series data; inferring the multi-source feature sequence using a deep learning model to obtain reservoir operational status assessment results and trend prediction results; the deep learning model consists of convolutional neural network branches, recurrent neural network branches, and attention mechanism branches connected in parallel; and determining the reservoir's operational status based on the reservoir operational status assessment results and trend prediction results. This application, used in reservoir operational status monitoring, solves the technical problem of low accuracy in existing reservoir operational status monitoring methods.

Description

Technical Field

[0001] This application relates to the field of water conservancy project operation monitoring, and in particular to a method and system for monitoring the operation status of reservoirs based on deep learning. Background Technology

[0002] Reservoirs, as crucial water infrastructure, fulfill multiple functions including flood control, water supply, power generation, and ecological regulation. Their operational status directly impacts flood control safety and water resource utilization efficiency in downstream areas. Current reservoir operation monitoring primarily relies on indicators such as water level, rainfall, and dam deformation, employing threshold discrimination or empirical rule-based analysis methods to assess operational status. However, these methods struggle to depict the complex relationships between hydrological conditions, engineering parameters, and environmental factors, and have limited predictive capabilities for operational status evolution. While advancements in sensor technology have resulted in the accumulation of substantial multi-source monitoring data during reservoir operation, existing technologies still fall short in areas such as joint modeling of multi-source data, feature interaction, and dynamic risk assessment, leading to issues like delayed status assessments and low early warning accuracy. Therefore, a new monitoring method is urgently needed to address the technical problem of low accuracy in current reservoir operation status monitoring. Summary of the Invention

[0003] This application provides a method and system for monitoring the operational status of reservoirs based on deep learning, which solves the technical problem of low accuracy in existing reservoir operational status monitoring.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a deep learning-based method for monitoring reservoir operation status is provided, comprising: collecting multi-source time-series data on reservoir operation, including: hydrological time-series data, engineering time-series data, and environmental time-series data; wherein, the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; and the environmental time-series data includes: water quality data and meteorological data; constructing a multi-source feature sequence of reservoir operation status based on the multi-source time-series data; inferring the multi-source feature sequence through a deep learning model to obtain reservoir operation status assessment results and trend prediction results; the deep learning model is composed of convolutional neural network branches, recurrent neural network branches, and attention mechanism branches connected in parallel; and determining the reservoir operation status based on the reservoir operation status assessment results and trend prediction results.

[0005] In conjunction with the first aspect mentioned above, one possible implementation involves constructing a multi-source feature sequence of reservoir operation status based on multi-source time-series data. This includes: dividing the hydrological time-series data into time windows, extracting water level change rate, cumulative rainfall, and seepage change within each time window to form a hydrological feature sequence; spatially grouping and performing temporal statistics on the engineering time-series data, calculating the displacement changes and trends at each monitoring location of the dam to form an engineering feature sequence; performing temporal statistical analysis on the environmental time-series data, extracting environmental state features reflecting changes in the reservoir's operating environment to form an environmental feature sequence; calculating the temporal correlation and time lag relationships between different feature sequences based on the hydrological feature sequence, engineering feature sequence, and environmental feature sequence to generate correlation features characterizing the mutual influence relationships of multi-source time-series data; and performing unified scale transformation and numerical normalization on the hydrological feature sequence, engineering feature sequence, environmental feature sequence, and correlation features to obtain the multi-source feature sequence.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves calculating the temporal correlation and time lag relationship between different feature sequences based on hydrological feature sequences, engineering feature sequences, and environmental feature sequences, thereby generating correlation features characterizing the mutual influence of multi-source time series data. This includes: aligning the hydrological feature sequences, engineering feature sequences, and environmental feature sequences along their time axes using a preset time resolution as a benchmark to construct a joint feature sequence set with a unified time index; setting multiple candidate time lag intervals within a preset time range, and calculating the lag feature between two different feature elements in the joint feature sequence set within each candidate time lag interval. Multiple pairs of lagged feature sequences are obtained from multiple candidate time lag intervals. Based on each pair of lagged feature sequences, a consistency index is calculated within the corresponding time window. The consistency index is used to characterize the correlation between feature sequences under different time lag conditions. Among multiple candidate time lag intervals, the candidate time lag interval with the optimal consistency index is selected, and the selected candidate time lag interval is determined as the time lag relationship between the corresponding feature sequences. Based on the time lag relationship and the corresponding consistency index, a correlation feature is constructed to characterize the mutual influence relationship between hydrological features, engineering features, and environmental features.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the change consistency index is obtained by weighting the trend consistency index, the magnitude correlation index, and the rate of change matching index. Among them, the trend consistency index Satisfy the following formula:

[0008] in, This represents the change in the characteristic sequence a at adjacent time points t. , This represents the eigenvalue of the feature sequence a at time point t. Let a represent the feature value of the feature sequence a at time point t-1. a and b are used to characterize any different class of water condition features, engineering condition features, or environmental features, respectively. This indicates that the lagged characteristic sequence b at time point The change at adjacent time points, , This indicates that the feature sequence b at time point eigenvalues, This indicates that the feature sequence b at time point eigenvalues, This refers to the time lag. This indicates an indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise; N is the number of valid time sample points participating in the calculation within the corresponding time window; In conjunction with the first aspect mentioned above, in one possible implementation, the correlation index... Satisfy the following formula:

[0009] in, This represents the eigenvalue of the feature sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. denoted by Pearson correlation coefficient function, where a and b are used to characterize any different class of water condition characteristics, engineering condition characteristics, or environmental characteristics, respectively; In conjunction with the first aspect mentioned above, in one possible implementation, the rate of change matching index Satisfy the following formula:

[0010] in, This represents the eigenvalue of the feature sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. This represents the rate of change of the characteristic sequence a at time point t. This indicates that the feature sequence b at time point The rate of change, This represents an exponential function.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, a deep learning model is used to infer multi-source feature sequences to obtain reservoir operation status assessment results and trend prediction results. This includes: inputting the multi-source feature sequences into the feature extraction branches corresponding to convolutional neural networks, recurrent neural networks, and attention mechanisms for parallel processing, including: in the convolutional neural network branch, using one-dimensional convolution operations to perform sliding convolution processing on the multi-source feature sequences to extract local change patterns of the feature sequences within different time windows, obtaining local temporal feature representations; in the recurrent neural network branch, modeling the evolution relationship of the feature sequences over a long time scale, obtaining global temporal feature representations; in the attention mechanism branch, calculating the importance weights of different time steps and different feature dimensions based on the multi-source feature sequences and their associated features, and weighting the corresponding features to obtain attention-enhanced feature representations; fusing the local temporal feature representations, global temporal feature representations, and attention-enhanced feature representations to obtain fused feature representations; and based on the fused feature representations, outputting the current reservoir operation status assessment results and the prediction results of the operation status change trends within a preset future time range through fully connected layers.

[0012] In conjunction with the first aspect mentioned above, one possible implementation involves fusing local temporal feature representations, global temporal feature representations, and attention-enhanced feature representations to obtain a fused feature representation. This includes: constructing a feature importance evaluation matrix based on the local temporal feature representations, global temporal feature representations, and attention-enhanced feature representations; dynamically generating fusion weights for each type of feature representation according to the feature importance evaluation matrix; and performing weighted fusion of the local temporal feature representations, global temporal feature representations, and attention-enhanced feature representations based on the fusion weights to obtain the fused feature representation.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the importance evaluation matrix satisfies the following formula:

[0014] in, Represents local temporal features; Represents global temporal features; This represents attention-enhancing feature representation; This is a trainable weight parameter matrix; Indicates feature concatenation operation; This represents element-wise multiplication and is used to characterize the interaction between different feature representations; This represents the normalization function.

[0015] In conjunction with the first aspect mentioned above, in one possible implementation, the reservoir's operational status is determined based on the reservoir's operational status assessment results and trend prediction results. This includes: standardizing the reservoir's operational status assessment results to obtain a current status risk index; applying time decay weighting to the trend prediction results to obtain a trend risk index characterizing the direction and intensity of changes in the reservoir's operational status; weighting and fusing the current status risk index and the trend risk index to obtain a comprehensive reservoir operational risk index; and comparing the comprehensive reservoir operational risk index with a preset risk level threshold to determine the corresponding reservoir operational status level.

[0016] Secondly, a deep learning-based reservoir operation status monitoring system is provided, comprising: a data acquisition device and an electronic device; wherein, the data acquisition device is used to collect multi-source time-series data of reservoir operation, including: hydrological time-series data, engineering time-series data, and environmental time-series data; wherein, the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; the environmental time-series data includes: water quality data and meteorological data; the electronic device is used to construct a multi-source feature sequence of reservoir operation status based on the multi-source time-series data; and to obtain reservoir operation status assessment results and trend prediction results by reasoning on the multi-source feature sequence through a deep learning model; the deep learning model is composed of convolutional neural network branches, recurrent neural network branches, and attention mechanism branches connected in parallel; and the reservoir operation status is determined based on the reservoir operation status assessment results and trend prediction results.

[0017] In conjunction with the second aspect mentioned above, in one possible implementation, the electronic device is also used to: divide the hydrological time series data into time windows, extract the water level change rate, cumulative rainfall, and seepage change within each time window to form a hydrological feature sequence; perform spatial grouping and temporal statistics on the engineering time series data, calculate the displacement change and trend at each monitoring location of the dam, and form an engineering feature sequence; perform temporal statistical analysis on the environmental time series data, extract environmental state characteristics reflecting changes in the reservoir's operating environment, and form an environmental feature sequence; based on the hydrological feature sequence, engineering feature sequence, and environmental feature sequence, calculate the temporal correlation and time lag relationship between different feature sequences, and generate correlation features characterizing the mutual influence relationship of multi-source time series data; and perform unified scale transformation and numerical normalization on the hydrological feature sequence, engineering feature sequence, environmental feature sequence, and correlation features to obtain a multi-source feature sequence.

[0018] This application provides a deep learning-based method and system for monitoring reservoir operation status. By uniformly modeling and collaboratively analyzing multi-source time-series data on reservoir water conditions, engineering conditions, and the environment, it introduces temporal correlation and time lag relationships between feature sequences, effectively characterizing the dynamic coupling mechanism among key factors during reservoir operation. Simultaneously, it employs a deep learning model structure combining convolutional neural networks, recurrent neural networks, and attention mechanisms in parallel. This allows for the extraction of local change features, long-term evolution features, and key feature weights at different time scales. Furthermore, a dynamic feature fusion mechanism adaptively adjusts the contribution of various features, thereby improving the accuracy and stability of reservoir operation status assessment and trend prediction. Further, by combining current status assessment results with future trend prediction results, a comprehensive operational risk index is constructed, enabling quantitative grading and forward-looking judgment of reservoir operation status. This enhances the comprehensiveness and reliability of reservoir operation monitoring and solves the technical problem of low accuracy in existing reservoir operation status monitoring.

[0019] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0020] Figure 1 A system architecture diagram of a reservoir operation status monitoring system based on deep learning is provided for embodiments of this application; Figure 2 A flowchart illustrating a deep learning-based method for monitoring the operational status of a reservoir, provided as an embodiment of this application; Figure 3 A flowchart illustrating another deep learning-based method for monitoring the operational status of a reservoir, provided as an embodiment of this application; Figure 4 A flowchart illustrating another deep learning-based method for monitoring the operational status of a reservoir, provided as an embodiment of this application; Figure 5 A flowchart illustrating another deep learning-based method for monitoring the operational status of a reservoir, provided in this application embodiment. Detailed Implementation

[0021] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0022] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] The deep learning-based reservoir operation status monitoring method provided in this application can be applied to, for example... Figure 1 The deep learning-based reservoir operation status monitoring system shown below, such as Figure 1 As shown, the system includes: a data acquisition device 101 and an electronic device 102; The data acquisition device 101 is used to collect multi-source time-series data on reservoir operation, including: hydrological time-series data, engineering time-series data, and environmental time-series data; the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; the environmental time-series data includes: water quality data and meteorological data. Electronic device 102 is used to construct a multi-source feature sequence of reservoir operation status based on multi-source time series data; to infer the multi-source feature sequence through a deep learning model to obtain reservoir operation status assessment results and trend prediction results; the deep learning model is composed of a convolutional neural network branch, a recurrent neural network branch and an attention mechanism branch connected in parallel; and to determine the reservoir operation status based on the reservoir operation status assessment results and trend prediction results.

[0024] In one possible implementation, the electronic device 102 is further configured to: divide the hydrological time series data into time windows, extract the water level change rate, cumulative rainfall, and seepage change within each time window to form a hydrological feature sequence; perform spatial grouping and temporal statistics on the engineering time series data, calculate the displacement change and trend at each monitoring location of the dam to form an engineering feature sequence; perform temporal statistical analysis on the environmental time series data, extract environmental state features reflecting changes in the reservoir's operating environment to form an environmental feature sequence; based on the hydrological feature sequence, engineering feature sequence, and environmental feature sequence, calculate the temporal correlation and time lag relationship between different feature sequences to generate correlation features characterizing the mutual influence relationship of multi-source time series data; and perform unified scale transformation and numerical normalization on the hydrological feature sequence, engineering feature sequence, environmental feature sequence, and correlation features to obtain a multi-source feature sequence.

[0025] To address the technical problem of low accuracy in existing reservoir operation status monitoring, this application provides a deep learning-based method for reservoir operation status monitoring. The method includes: collecting multi-source time-series data on reservoir operation, including: hydrological time-series data, engineering time-series data, and environmental time-series data; wherein the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; and the environmental time-series data includes: water quality data and meteorological data; constructing a multi-source feature sequence of reservoir operation status based on the multi-source time-series data; inferring the multi-source feature sequence using a deep learning model to obtain reservoir operation status assessment results and trend prediction results; the deep learning model consists of convolutional neural network branches, recurrent neural network branches, and attention mechanism branches connected in parallel; and determining the reservoir operation status based on the reservoir operation status assessment results and trend prediction results.

[0026] Figure 2 A flowchart illustrating the deep learning-based reservoir operation status monitoring method provided in this application embodiment is shown below. Figure 2 As shown, the method includes: S201. Collect multi-source time-series data on reservoir operation.

[0027] Multi-source time-series data refers to various types of monitoring data continuously collected at a unified or alignable time scale, reflecting changes in the reservoir's operational status. This includes hydrological time-series data, engineering time-series data, and environmental time-series data. Hydrological time-series data characterizes the reservoir's hydrological changes, engineering time-series data characterizes the operation and structural response characteristics of hydraulic structures, and environmental time-series data characterizes changes in external natural environmental conditions. Hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; engineering time-series data includes: dam deformation time-series data and dam structural performance data; environmental time-series data includes: water quality data and meteorological data.

[0028] In one possible implementation, various monitoring indicators are continuously or periodically sampled using water level gauges, rain gauges, piezometers, displacement sensors, water quality sensors, and meteorological monitoring equipment deployed at the reservoir site. The collected raw data is marked with a unified timestamp and transmitted to the data processing unit via wired or wireless communication, thereby forming a multi-source, multi-dimensional raw time-series data set with temporal order.

[0029] It should be noted that different types of monitoring equipment can have different sampling frequencies. Frequency uniformity is not mandatory during the data acquisition phase. It is only necessary to ensure that the data has a traceable time identifier so that time alignment and joint analysis can be performed in subsequent processing. At the same time, it is permissible for some monitoring data to have missing or outlier values.

[0030] S202. Based on multi-source time series data, construct a multi-source feature sequence of reservoir operation status.

[0031] Among them, multi-source feature sequences refer to feature expression sequences obtained from multi-source time-series data after feature extraction and formal representation, which can be used as inputs to deep learning models and are used to uniformly describe the changing characteristics of different types of monitoring information during reservoir operation.

[0032] In one possible implementation, the acquired multi-source time-series data is segmented according to a preset time window, and statistical calculations, change calculations, or state quantity extractions are performed on various types of data within each time window to convert the original monitoring data into feature vectors with clear physical meaning and numerical stability; features from different sources are combined in chronological order to form a multi-source feature sequence.

[0033] It should be noted that in the process of constructing multi-source feature sequences, the features can be scaled and normalized according to the differences in the dimensions and numerical ranges of the monitoring data, so as to avoid the excessive numerical differences between different features from adversely affecting the model training and inference process.

[0034] Based on the above steps, this step improves the standardization and stability of data representation by transforming the original monitoring data into a structured multi-source feature sequence, enabling subsequent deep learning models to perform feature learning and inference efficiently and reliably.

[0035] S203. By using a deep learning model to infer multi-source feature sequences, the reservoir operation status assessment results and trend prediction results are obtained.

[0036] In this application, the deep learning model adopts a multi-branch parallel structure to adapt to the differences in time scale, change patterns, and influencing mechanisms of multi-source time-series data on reservoir operation. The deep learning model consists of a convolutional neural network branch, a recurrent neural network branch, and an attention mechanism branch connected in parallel. The convolutional neural network branch performs one-dimensional convolution operations on the multi-source feature sequences, extracting local fluctuation patterns and abrupt changes in features such as water level, rainfall, seepage, and structural response within a short time window using convolutional kernels of different scales, thereby characterizing the instantaneous changes in the reservoir's operating state. The recurrent neural network branch performs time-series modeling of the multi-source feature sequences, accumulating and updating long-term historical information through memory units, characterizing the evolutionary patterns and cumulative effects of the reservoir's operating state over a longer time scale. The attention mechanism branch dynamically allocates weights for different time steps and feature dimensions based on the multi-source feature sequences and their associated features, enabling the model to adaptively focus on key factors that significantly influence the current operating state and future trends. The feature representations output by the three branches running in parallel are fused at the feature layer to form a comprehensive representation of the reservoir's operating state. The reservoir operation status assessment results are used to quantitatively reflect the reservoir's operational safety or health level at the current moment, while the trend prediction results are used to depict the possible direction and intensity of changes in the reservoir's operation status within a preset time range in the future, thereby providing a basis for reservoir operation management and risk early warning.

[0037] In one possible implementation, multi-source feature sequences are simultaneously input into a deep learning model consisting of convolutional neural network branches, recurrent neural network branches, and attention mechanism branches connected in parallel. Each branch processes the input features in parallel. After completing feature extraction and fusion, the model generates an evaluation result of the current running state and a prediction result of the running state change trend within a preset time range through the output layer.

[0038] It should be noted that deep learning models can be trained offline based on historical monitoring data and corresponding known operating status samples, and then used for online inference after training. At the same time, the model parameters can be periodically updated based on newly collected data to adapt to long-term changes in reservoir operating conditions.

[0039] As an example, historical operational data can be used to train a model that outputs a safety level score corresponding to the current reservoir operation status, and simultaneously predicts the trend of this score over several future time steps.

[0040] Based on the above steps, this step uses a deep learning model to perform joint reasoning on multi-source feature sequences, thereby achieving automated assessment and trend prediction of reservoir operation status and reducing reliance on human experience judgment.

[0041] S204. Based on the reservoir operation status assessment results and trend prediction results, determine the reservoir operation status.

[0042] Among them, the reservoir operation status refers to the reservoir operation level or risk status determined by a comprehensive assessment of current status information and future trend information, which is used to support operation management and risk early warning decision-making.

[0043] In one possible implementation, the reservoir operation status assessment results are standardized to obtain the current status risk index; the trend prediction results are subjected to time decay weighting to obtain the trend risk index characterizing the direction and intensity of changes in the reservoir operation status; the current status risk index and the trend risk index are weighted and fused to obtain the comprehensive reservoir operation risk index; the comprehensive reservoir operation risk index is compared with the preset risk level threshold to determine the corresponding reservoir operation status level.

[0044] It should be noted that the weights of the status indicators and trend indicators, as well as the risk level thresholds, can be set and adjusted according to the reservoir size, engineering level, or management needs, and are not limited to fixed values. In this application, the reservoir operating status is a quantitative and hierarchical representation of the overall operating condition of the reservoir at a specific moment and over a future period. Specifically, it is reflected in the reservoir's status level in terms of water volume regulation, engineering structural safety, and adaptability to the operating environment. It not only reflects whether the reservoir is currently operating safely and stably, but also reflects the evolution direction of potential risks. Specifically, the reservoir's operational status includes a comprehensive assessment of at least one or more of the following status types: The reservoir's operational status can be characterized by preset levels such as normal operation, monitoring operation, early warning operation, and high-risk operation. Normal operation indicates that the reservoir's water level, seepage, dam deformation, and environmental indicators are all within the design limits, and trend predictions show no significant risk accumulation in the short term. Monitoring operation indicates that some indicators are approaching the early warning threshold or showing a slow, abnormal trend, requiring enhanced monitoring and scheduling. Early warning operation indicates that multiple key indicators are abnormal or trend predictions show an amplifying risk, requiring scheduling or engineering intervention measures. High-risk operation indicates that the reservoir's operating parameters significantly deviate from the safe range or trend predictions show an instability risk, requiring immediate emergency response. Furthermore, the reservoir's operational status can be quantitatively characterized by a continuous comprehensive operational risk index. This index is weighted by current status risk indicators and trend risk indicators, reflecting the magnitude and direction of change of reservoir operational risk, thus providing clear and actionable status information for reservoir scheduling decisions, risk early warning, and operation and maintenance management.

[0045] As an example, the comprehensive operational risk index R of the reservoir is set as a normalized continuous value, ranging from 0 to 1. Multiple risk level thresholds are preset according to the needs of engineering operation safety management. When R < 0.25R, the reservoir is determined to be in normal operation; when 0.25 ≤ R < 0.50, the reservoir is determined to be in a state of concern; when 0.50 ≤ R < 0.75, the reservoir is determined to be in a state of early warning; and when R ≥ 0.75, the reservoir is determined to be in a state of high risk. The risk level thresholds can be adjusted based on the reservoir's engineering scale, dam structure, historical operational data, and management experience to achieve graded judgment and dynamic control of the reservoir's operational status.

[0046] This application's embodiments, through unified acquisition and modeling of multi-source time-series data, can simultaneously reflect reservoir water level changes, engineering structure responses, and external environmental influences, avoiding information gaps or judgment biases caused by single data sources. By constructing multi-source feature sequences and introducing parallel deep learning models, features of different time scales and mechanisms are specifically modeled and fused, effectively improving the accuracy of identifying the current operating status of the reservoir and the ability to predict the evolution trend of potential risks. Finally, based on the state assessment results and trend prediction results, the reservoir operating status is comprehensively determined, so that the operating status determination not only reflects "whether it is safe now" but also "whether there is a risk of deterioration in the future," thereby enhancing the foresight and reliability of reservoir operation management, reducing the risk of sudden risks and misjudgments, realizing continuous, comprehensive, and refined monitoring of the reservoir operating status, and solving the technical problem of low accuracy in existing reservoir operating status monitoring.

[0047] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S202 can be specifically implemented through the following S301 to S305, which are explained in detail below: S301. Divide the hydrological time series data into time windows, and extract the water level change rate, cumulative rainfall and seepage change within each time window to form a hydrological characteristic sequence.

[0048] The time window refers to the time interval in which continuous time-series data is segmented and processed with a fixed time length.

[0049] In one possible implementation, the time window length is set according to the sampling frequency of the monitoring data, for example, 1 hour, 3 hours, or 24 hours as a time window. For the water level time series data, the difference in water level between adjacent moments is calculated within each time window and divided by the time interval to obtain the water level change rate. For the rainfall time series data, the cumulative rainfall is calculated within the corresponding time window. For the seepage time series data, the difference between the start and end times is calculated within the time window to obtain the seepage change. Finally, the water level change rate, cumulative rainfall, and seepage change obtained in each time window are combined in chronological order to form a hydrological characteristic sequence.

[0050] As an example, in the flood season operation scenario, a 1-hour time window can be used to calculate the water level change rate from hourly water level data, accumulate the hourly rainfall from rain gauges, and simultaneously calculate the change in seepage monitoring points on the dam body within that hour, thereby forming a water condition feature sequence that reflects the characteristics of short-term water condition changes.

[0051] Based on the above steps, this step can transform the original high-frequency, noisy hydrological monitoring data into a hydrological feature sequence with clear physical meaning and time scale characteristics. This helps to highlight the key trends in water level, rainfall, and seepage changes and reduce the interference of original data fluctuations on subsequent model inference.

[0052] S302. Perform spatial grouping and temporal statistics on the time series data of the dam condition, calculate the displacement changes and trends at each monitoring location of the dam, and form a characteristic sequence of the dam condition.

[0053] Spatial grouping refers to grouping monitoring points according to their spatial location or structural position on the dam.

[0054] In one possible implementation, based on the dam structure layout, deformation monitoring points are spatially grouped according to dam sections, dam axis directions, or elevation intervals. Within each spatial group, time-series displacement data of each monitoring point are statistically analyzed to calculate the displacement change within the corresponding time window. Furthermore, by performing linear fitting or difference analysis on the displacement changes of multiple consecutive time windows, the displacement change trend is extracted. Finally, the displacement changes and trends of each spatial group within each time window are combined to form a sequence of engineering condition characteristics.

[0055] It should be noted that different types of dams can adopt different spatial grouping strategies, and structural performance data can be used as auxiliary features and incorporated into the corresponding spatial grouping to enhance the ability to characterize the structural operating status.

[0056] As an example, for a concrete dam with multiple vertical measuring lines, the measuring points can be grouped according to the dam section, and the horizontal displacement change and its trend of each dam section can be calculated within each time window, thereby forming a sequence of engineering characteristics that reflects the overall and local deformation behavior of the dam body.

[0057] Based on the above steps, this step can highlight the characteristics of dam deformation evolution over time while maintaining structural spatial information, so that the engineering condition feature sequence has both engineering physical significance and is convenient for subsequent deep learning models to perform time series modeling.

[0058] S303. Perform time statistical analysis on environmental time series data, extract environmental state characteristics that reflect changes in the reservoir's operating environment, and form an environmental characteristic sequence.

[0059] Among them, environmental state characteristics are statistical features used to characterize the degree and trend of changes in environmental conditions.

[0060] In one possible implementation, environmental time-series data are statistically analyzed according to time windows consistent with the hydrological characteristic sequence. Mean, extreme values, or ranges of change within the time window are extracted from water quality data to reflect the state of the aquatic environment; statistics on elements such as temperature, air pressure, and wind speed are extracted from meteorological data to reflect external climate conditions; and the environmental state characteristics obtained within each time window are arranged in chronological order to form an environmental characteristic sequence.

[0061] It should be noted that the selection of environmental characteristics can be adjusted according to the needs of reservoir operation and management. For example, during the high-temperature season, the focus should be on water temperature changes, and under extreme weather conditions, the focus should be on rainfall and wind speed characteristics.

[0062] As an example, under summer operating conditions, the average water temperature, dissolved oxygen variation range, and daily air temperature variation range can be calculated within a 24-hour window to form an environmental characteristic sequence reflecting the reservoir's environmental status.

[0063] Based on the above steps, this step can integrate scattered environmental monitoring data into a stable and comparable environmental feature sequence, thereby enhancing the model's ability to perceive the impact of the external environment on reservoir operation.

[0064] S304. Based on hydrological characteristic sequences, engineering characteristic sequences, and environmental characteristic sequences, calculate the temporal correlation and time lag relationship between different characteristic sequences, and generate correlation features that characterize the mutual influence relationship of multi-source time series data.

[0065] Among them, time correlation is used to characterize the consistency of the direction and magnitude of changes in different feature sequences, while time lag is used to describe the time delay when a change in one type of feature affects another type of feature.

[0066] In one possible implementation, the hydrological feature sequence, engineering feature sequence, and environmental feature sequence are aligned along the time axis. Then, within a preset time lag range, the correlation index of any two types of feature sequences is calculated under different lag conditions. Based on the magnitude of the correlation index, the optimal time lag is selected, and the lag and the corresponding correlation result are output as the associated feature.

[0067] It should be noted that the time lag range can be limited based on reservoir engineering experience in order to avoid long time lags with no physical meaning from interfering with the effectiveness of the correlation characteristics.

[0068] As an example, the correlation between "cumulative rainfall" and "dam displacement change" within a lag range of 0 to 72 hours can be calculated, and the lag time with the highest correlation can be selected as the correlation feature between the two.

[0069] Based on the above steps, this step can explicitly characterize the mutual influence relationship between multi-source features, enabling the model to not only focus on the changes of a single feature, but also learn the temporal coupling behavior between different features.

[0070] S305. Perform unified scale transformation and numerical normalization on the water condition feature sequence, engineering condition feature sequence, environmental feature sequence, and related features to obtain a multi-source feature sequence.

[0071] Among them, unified scale conversion refers to mapping features with different dimensions and numerical ranges to a comparable numerical space, while numerical normalization refers to standardizing or normalizing feature values.

[0072] In one possible implementation, the statistical range of each type of feature is calculated, and the standard deviation normalization method is used to map the water condition features, engineering condition features, environmental features and related features to a unified numerical range; then, the normalized features are combined according to the time index to form a multi-source feature sequence with a unified structure.

[0073] As an example, features of different dimensions such as water level change rate, displacement change, and correlation index can be uniformly normalized to the [0,1] interval and combined into a multidimensional feature vector sequence in chronological order.

[0074] Based on the above steps, this step can effectively eliminate the differences in the dimensions and numerical ranges of different features, enabling multi-source features to have good learnability and numerical stability in subsequent deep learning models.

[0075] This application's embodiments, through hierarchical feature extraction, correlation modeling, and unified normalization of multi-source time-series data including hydrological, engineering, and environmental data, can transform raw, heterogeneous, and significantly scale-differentiated monitoring data into structurally consistent, time-aligned, and semantically clear multi-source feature sequences. On the one hand, it effectively preserves dynamic information that has a critical impact on reservoir operation safety, such as water level changes, dam deformation, and environmental disturbances. On the other hand, by introducing correlation features based on time correlation and time lag, it explicitly characterizes the interaction mechanisms between different operational elements, thereby reducing information redundancy and noise interference, and improving the discriminativeness and interpretability of feature expressions. Simultaneously, unified scale transformation and numerical normalization ensure the comparability and stability of features from different sources in subsequent deep learning models, avoiding model training bias caused by differences in dimensions. This makes the constructed multi-source feature sequences more suitable for reservoir operation status assessment and trend prediction, which is conducive to improving the accuracy, robustness, and applicability of the overall analysis results.

[0076] In one possible implementation of the embodiments of this application, combined with Figure 3 ,like Figure 4As shown, the above S304 can be specifically implemented through the following S401 to S405, which are explained in detail below: S401. Based on the preset time resolution, align the water condition feature sequence, engineering condition feature sequence, and environmental feature sequence with time axis to construct a set of joint feature sequences with a unified time index.

[0077] Among them, the preset time resolution refers to the minimum time interval used to unify the time index of multi-source feature sequences, such as 5 minutes, 10 minutes or 1 hour; time axis alignment refers to mapping feature sequences with different sampling frequencies and different timestamps to the same time base to eliminate the impact of time inconsistency.

[0078] In one possible implementation, a unified time resolution is determined as a common time axis, and the water condition feature sequence, engineering condition feature sequence, and environmental feature sequence are resampled according to this time resolution. For high-frequency sampled data, the statistical feature values ​​within the corresponding time window are calculated by time aggregation. For low-frequency or missing data, interpolation or forward padding is used to complete the data, so that the three types of features form a joint feature sequence set under the same time index.

[0079] It should be noted that the fidelity of the original physical meaning must be ensured during the timeline alignment process to avoid introducing spurious changes due to excessive interpolation or unreasonable aggregation. At the same time, different types of features can adopt differentiated resampling strategies to balance data integrity and temporal continuity.

[0080] As an example, when water level and rainfall data are sampled every 5 minutes, while dam displacement data are sampled every 30 minutes, 10 minutes can be selected as a unified time resolution. Window statistics can be performed on the hydrological data, and linear interpolation can be used to complete the engineering data, thereby generating a joint feature sequence with a unified time index.

[0081] Based on the above steps, this step can effectively eliminate the differences in sampling frequency and timestamps of multi-source monitoring data, providing a consistent and comparable time basis for subsequent time lag analysis and reducing the impact of time series misalignment on the accuracy of correlation modeling.

[0082] S402. Set multiple candidate time lag intervals within a preset time range, and calculate the lag feature between two different feature elements in the joint feature sequence set in each candidate time lag interval to obtain multiple sets of lag feature sequence pairs for multiple candidate time lag intervals.

[0083] The preset time range refers to the time delay search interval pre-set based on the reservoir's hydrological response characteristics and engineering operation experience when analyzing the time lag relationship between different characteristic sequences. It is used to limit the reasonable range of time lag analysis. The preset time range is usually set in hours or days, and its upper and lower limits can be determined according to the reservoir's catchment area, storage capacity, dam structure type, and monitoring data sampling frequency. For example, it can be set to 0-6 hours, 0-12 hours, or 0-24 hours. The candidate time lag interval refers to multiple time ranges that may have causal or response delays, such as 0-1 hour, 1-3 hours, and 3-6 hours. The lag feature refers to the new feature sequence formed by shifting a certain feature sequence forward or backward on the time axis.

[0084] In one possible implementation, for any two different types of feature elements in the joint feature sequence set, a time shift operation is performed on one type of feature sequence within each candidate time lag interval, and it is combined with the other type of feature sequence to form a lag feature sequence pair within the same time window, thereby constructing multiple sets of candidate lag feature sequence pairs under different lag conditions.

[0085] It should be noted that the setting of candidate time lag intervals should be based on reservoir engineering experience and historical operating patterns to avoid excessively large lag ranges that could dilute the correlation; at the same time, reasonable coverage and distinguishability should be maintained between each candidate interval.

[0086] As an example, when analyzing the relationship between rainfall and seepage changes in the dam body, lag characteristic sequences with rainfall lags of 1 hour, 3 hours, and 6 hours can be constructed respectively, and these sequences can be combined with the seepage characteristics at the current moment to form multiple lag characteristic sequence pairs.

[0087] Based on the above steps, this step can systematically enumerate the characteristic combination relationships under different time lag assumptions, enabling the potential delayed response characteristics to be explicitly modeled, thereby improving the comprehensiveness and reliability of the correlation analysis.

[0088] S403. Calculate the consistency index of change within the corresponding time window for each of the multiple sets of lagged feature sequence pairs.

[0089] Among them, the consistency index is used to characterize the degree of correlation between feature sequences under different time lag conditions, including trend consistency, magnitude correlation and rate of change matching.

[0090] In one possible implementation, for each pair of lagged feature sequences, the changes in the two features are calculated within the corresponding time window, and their consistency scores are calculated using statistical methods, thereby obtaining the value of the change consistency index under the time lag condition.

[0091] As an example, the trend consistency indicator Satisfy the following formula:

[0092] in, This represents the change in the characteristic sequence a at adjacent time points t. , This represents the eigenvalue of the feature sequence a at time point t. Let a represent the feature value of the feature sequence a at time point t-1. a and b are used to characterize any different class of water condition features, engineering condition features, or environmental features, respectively. This indicates that the lagged characteristic sequence b at time point The change at adjacent time points, , This indicates that the feature sequence b at time point eigenvalues, This indicates that the feature sequence b at time point eigenvalues, This refers to the time lag. This indicates an indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise; N is the number of valid time sample points participating in the calculation within the corresponding time window; As an example, correlation metrics Satisfy the following formula:

[0093] in, This represents the eigenvalue of the feature sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. denoted by Pearson correlation coefficient function, where a and b are used to characterize any different class of water condition characteristics, engineering condition characteristics, or environmental characteristics, respectively; As an example, the rate of change matching index Satisfy the following formula:

[0094] in, This represents the eigenvalue of the feature sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. This represents the rate of change of the characteristic sequence a at time point t. This indicates that the feature sequence b at time point The rate of change, This represents an exponential function.

[0095] Based on the above steps, this step can quantitatively reflect the degree of coordinated change of different features under different time lag conditions, providing an objective evaluation basis for identifying the most representative time-series influence relationships.

[0096] S404. Among multiple candidate time lag intervals, select the candidate time lag interval with the optimal change consistency index, and determine the selected candidate time lag interval as the time lag relationship between the corresponding feature sequences.

[0097] Among them, the optimal time lag relationship refers to the time lag setting that makes the consistency index of change reach the maximum or optimal level among multiple candidate time lag intervals.

[0098] In one possible implementation, the consistency index of changes corresponding to each candidate time lag interval is compared, the time lag interval with the best index value is selected, and the time offset corresponding to the interval is determined as the time lag relationship between the corresponding feature sequences.

[0099] It should be noted that when the consistency index of changes in multiple candidate intervals is close, stability or continuity constraints can be introduced for secondary screening to avoid frequent jumps in time lag relationships.

[0100] As an example, if the consistency index of change corresponding to a rainfall lag of 3 hours is significantly higher than that of 1 hour and 6 hours, then 3 hours is determined to be the optimal time lag relationship between rainfall and seepage.

[0101] Based on the above steps, this step can automatically identify the most representative time-series response relationship among multiple source features, making the time lag modeling results more consistent with the actual engineering operation rules.

[0102] S405. Based on the time lag relationship and the corresponding change consistency index, construct correlation features to characterize the mutual influence between water condition features, engineering condition features and environmental features.

[0103] Among them, the correlation feature refers to the high-level semantic feature formed by integrating time lag relationship and change consistency index, which is used to characterize the intensity and direction of mutual influence between different operational elements.

[0104] In one possible implementation, the lagged feature values ​​corresponding to the determined time lag relationship are combined and encoded with the change consistency index to generate associated features that can reflect the direction, intensity and time delay characteristics of the feature influence, and these features are incorporated into the multi-source feature sequence as a new feature dimension.

[0105] Based on the above steps, this step can explicitly introduce the mutual influence information between multiple operational elements, so that subsequent deep learning models not only rely on changes in a single feature, but also utilize cross-feature and cross-time correlations, thereby improving the refinement and interpretability of reservoir operation status analysis.

[0106] This application introduces an explicit time alignment and time lag modeling mechanism during the construction of multi-source time-series features, enabling a fine-grained characterization of the causal response relationships between hydrological features, engineering features, and environmental features in the time dimension. This avoids the correlation distortion problem caused by neglecting physical response delays in traditional synchronous modeling methods. By constructing multiple sets of lag features within a preset time range and introducing a change consistency index for quantitative screening, the application can adaptively identify the most representative time lag relationships between different features, thereby improving the expressive power of associated features for the actual operating mechanism. Furthermore, by constructing associated features based on the optimal time lag relationship and corresponding consistency index, the application effectively enhances the structural coupling and information complementarity between multi-source features, enabling subsequent deep learning models to more accurately capture the dynamic evolution characteristics of reservoir operation status during state assessment and trend prediction, thereby improving the reliability and stability of the overall monitoring results.

[0107] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, S203 above uses a deep learning model to infer multi-source feature sequences to obtain reservoir operation status assessment results and trend prediction results. This includes: inputting the multi-source feature sequences into the feature extraction branches corresponding to the convolutional neural network, recurrent neural network, and attention mechanism for parallel processing. Specifically, this can be achieved through S501 to S505, which are explained in detail below: S501. In the convolutional neural network branch, a one-dimensional convolution operation is used to perform sliding convolution processing on multi-source feature sequences to extract the local change patterns of the feature sequences in different time windows and obtain local temporal feature representations.

[0108] Local temporal feature representation refers to the feature representation that characterizes the magnitude of change, mutation pattern and local trend of multi-source feature sequences on a short time scale through convolution operations within a finite time window; one-dimensional convolution is a convolution operation that slides along the time dimension to calculate the feature sequence, and is suitable for local pattern extraction of time series data.

[0109] In one possible implementation, the multi-source feature sequence after uniform scaling and normalization is input into the convolutional neural network branch according to the time dimension. Multiple convolutional kernels of different sizes are set to perform one-dimensional sliding convolution operation on the feature sequence. Each convolutional kernel performs weighted summation of feature values ​​within the corresponding time window and processes them through a nonlinear activation function, thereby extracting local temporal features that reflect changes in water level, rainfall fluctuations, sudden changes in seepage, and structural response changes in a short period of time, and finally forming a local temporal feature representation.

[0110] It should be noted that the time window length of the convolution kernel can be set according to the data sampling frequency of reservoir operation monitoring to adapt to the local change characteristics at different time scales; at the same time, multiple local change patterns can be extracted in parallel through multi-channel convolution to avoid the problem that a single time scale is insufficient to represent complex operational features.

[0111] As an example, in a data scenario with a sampling interval of 5 minutes, convolutional kernels of lengths 3, 6, and 12 can be set to correspond to local change features within 15 minutes, 30 minutes, and 1 hour, respectively, to characterize rapid changes in water level or local seepage anomalies caused by short-term heavy rainfall.

[0112] Based on the above steps, this step explicitly extracts the local variation patterns of multi-source features within a short time scale through convolution, enabling the model to sensitively identify sudden or phased changes during reservoir operation and improve its ability to perceive abnormal operating conditions and rapid evolution states.

[0113] S502. In the recurrent neural network branch, the evolution of feature sequences over a long time scale is modeled to obtain a global temporal feature representation.

[0114] Among them, global temporal feature representation refers to the feature representation that reflects the continuous evolution of reservoir operation status over time by modeling the dependency relationship between feature sequences over a long period of time; recurrent neural networks are used to process time-dependent sequence data.

[0115] In one possible implementation, multi-source feature sequences are input into a branch of a recurrent neural network in chronological order. The recurrent neural network, through a recursive update mechanism of hidden states, fuses the feature input of the current time step with the state information of historical time steps, continuously accumulating long-term evolution information of water conditions, engineering conditions, and environmental characteristics, and outputs a global temporal feature representation that reflects the overall trend and long-term change law of reservoir operation.

[0116] It should be noted that recurrent neural networks can employ long short-term memory networks or gated recurrent units to alleviate the gradient vanishing problem in long-term sequence modeling, thereby stably capturing the changing trends of reservoir operation status over a longer period of time.

[0117] Based on the above steps, this step can characterize the overall change process of the reservoir's operating status from the perspective of time evolution, enabling the model to have the ability to perceive long-term trends and avoid the one-sidedness of state judgment caused by relying solely on short-term characteristics.

[0118] S503. In the attention mechanism branch, based on the multi-source feature sequence and its associated features, the importance weights of different time steps and different feature dimensions are calculated, and the corresponding features are weighted to obtain the attention-enhanced feature representation.

[0119] Among them, attention-enhanced feature representation refers to assigning differentiated weights to different time steps and different feature dimensions, so that the model focuses on feature representations that contribute more to the reservoir's operational status; correlation features are used to reflect the time lag and correlation between multi-source features.

[0120] In one possible implementation, multi-source feature sequences and their associated features are input into the attention mechanism branch. By constructing query vectors, key vectors, and value vectors, the correlation between different time steps and different feature dimensions is calculated, and attention weights are generated accordingly. Then, the original features are weighted and summed using these attention weights to obtain attention-enhanced feature representations that emphasize key time nodes and key feature dimensions.

[0121] It should be noted that the calculation of attention weights can be combined with the aforementioned time lag relationship, so that the model can fully consider the physical response delay factor when calculating attention, and avoid assigning too high weights to non-critical time points or redundant features.

[0122] Based on the above steps, this step dynamically allocates feature weights, enabling the model to focus on the information most discriminative of changes in reservoir operation status, thereby improving the relevance and effectiveness of feature representation.

[0123] S504. The local temporal feature representation, the global temporal feature representation, and the attention-enhanced feature representation are fused to obtain the fused feature representation.

[0124] Among them, fusion feature representation refers to the high-dimensional feature representation obtained by comprehensively expressing local temporal feature representation, global temporal feature representation and attention-enhanced feature representation.

[0125] In one possible implementation, a feature importance evaluation matrix is ​​constructed based on local temporal feature representation, global temporal feature representation, and attention-enhanced feature representation. According to the feature importance evaluation matrix, the fusion weights corresponding to each type of feature representation are dynamically generated, and the local temporal feature representation, global temporal feature representation, and attention-enhanced feature representation are weighted and fused according to the fusion weights to obtain the fused feature representation.

[0126] As an example, the sexual evaluation matrix satisfies the following formula:

[0127] in, Represents local temporal features; Represents global temporal features; This represents attention-enhancing feature representation; This is a trainable weight parameter matrix; Indicates feature concatenation operation; This represents element-wise multiplication and is used to characterize the interaction between different feature representations; This represents the normalization function.

[0128] Based on the above steps, this step achieves effective collaborative expression of multi-scale and multi-perspective feature information, enabling the fused features to comprehensively reflect the complex characteristics of the reservoir's operating status.

[0129] S505. Based on the fusion feature representation, the current reservoir operation status assessment results and the prediction results of the operation status change trend within a preset time range are output through the fully connected layer.

[0130] In one possible implementation, the fused feature representation is used as input and fed into two independent fully connected output layers for mapping calculation. One fully connected output layer is used to generate the current operation status assessment result of the reservoir, and the other fully connected output layer is used to generate the operation status change trend prediction result within a future preset time range. The parameters of each fully connected output layer are obtained by training with historical labeled samples, so that the output results can accurately reflect the operation status of the reservoir and its change trend.

[0131] It should be noted that the state assessment results and trend prediction results can be trained using different loss functions to balance the accuracy of current state judgment and the stability of future trend prediction.

[0132] As an example, the status assessment results can output the corresponding operational risk level, and the trend prediction results can output the risk index change curve for the next 6 or 24 hours.

[0133] Based on the above steps, this step enables real-time assessment and forward-looking prediction of the reservoir's operational status, so that the monitoring results not only reflect the current operational status, but also provide decision-making references for future risk evolution.

[0134] This application achieves refined assessment and trend prediction of reservoir operation status by collaboratively modeling multi-source time-series data and introducing a deep learning model. Compared with traditional methods that rely solely on a single model or single-scale features, this application can simultaneously characterize the local changes at short-term time scales, the overall evolution at long-term time scales, and the differentiated impacts of key time nodes and key feature dimensions during reservoir operation. This makes the operation status assessment results more comprehensive, stable, and discriminative. Furthermore, by integrating the status assessment results with the trend prediction results to comprehensively determine the reservoir operation status, it can not only reflect the current safety or health level of the reservoir but also reveal the direction and intensity of potential risks in advance. This improves the accuracy, foresight, and practicality of reservoir operation monitoring, providing more reliable technical support for operation management and risk early warning.

[0135] In one possible implementation, the training process of the deep learning model includes: collecting multi-source time-series data and corresponding operational status annotation information during the historical operation phase of the reservoir, including the reservoir operational status level and / or comprehensive operational risk index at the historical moment; performing sliding segmentation of the historical multi-source time-series data with a fixed time window length T, where the time window length T is set to 24h and the sliding step size between adjacent windows can be set to 1h; sequentially performing hydrological feature extraction, engineering feature extraction, environmental feature extraction, and association feature construction on the segmented multi-source time-series data, and normalizing the obtained multi-source feature sequences as input samples for the model. During model training, multi-source feature sequences are input into the convolutional neural network (CNN) branch, the recurrent neural network (RNN) branch, and the attention mechanism branch, respectively. The CNN branch uses a one-dimensional convolutional structure with 32 kernels and a kernel length of 3. The RNN branch uses an LSTM structure with a hidden layer dimension of 64. The attention mechanism branch uses a multi-head attention structure with 4 attention heads. After feature fusion of the output features from each branch, the reservoir operation status assessment results and trend prediction results are output through fully connected layers. The weighted sum of the operation status assessment error and the trend prediction error is used as the training loss function, with the weight coefficient of each sub-loss set to 0.4. The backpropagation algorithm is used to update the model parameters, with Adam as the optimizer and an initial learning rate of 1×10⁻⁶. - The training batch size can be set to 8, and the number of training rounds can be set to 50. Training is terminated when the validation set loss function converges or the preset number of training rounds is reached, thus obtaining a trained deep learning model for reservoir operation status monitoring.

[0136] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0137] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A method for monitoring the operational status of a reservoir based on deep learning, characterized in that, include: Multi-source time-series data on reservoir operation are collected, including: hydrological time-series data, engineering time-series data, and environmental time-series data; wherein, the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; the environmental time-series data includes: water quality data and meteorological data; Based on the multi-source time-series data, a multi-source feature sequence of reservoir operation status is constructed. The reservoir operation status assessment results and trend prediction results are obtained by reasoning on the multi-source feature sequence through a deep learning model; the deep learning model is composed of a convolutional neural network branch, a recurrent neural network branch and an attention mechanism branch connected in parallel. Based on the reservoir operation status assessment results and trend prediction results, the reservoir operation status is determined.

2. The method according to claim 1, characterized in that, The construction of a multi-source feature sequence of reservoir operation status based on the multi-source time-series data includes: The hydrological time series data is divided into time windows, and the water level change rate, cumulative rainfall and seepage change are extracted in each time window to form a hydrological feature sequence. Spatial grouping and temporal statistics are performed on the aforementioned engineering condition time series data to calculate the displacement changes and trends at each monitoring location of the dam, forming an engineering condition characteristic sequence; The environmental time series data are subjected to time statistical analysis to extract environmental state characteristics that reflect changes in the reservoir's operating environment, forming an environmental characteristic sequence; Based on the aforementioned hydrological characteristic sequence, engineering characteristic sequence, and environmental characteristic sequence, the temporal correlation and time lag relationship between different characteristic sequences are calculated to generate correlation features that characterize the mutual influence relationship of multi-source time series data. The hydrological feature sequence, engineering feature sequence, environmental feature sequence, and associated features are subjected to unified scale transformation and numerical normalization to obtain a multi-source feature sequence.

3. The method according to claim 2, characterized in that, The process of calculating the temporal correlation and time lag between different feature sequences based on the hydrological feature sequence, engineering feature sequence, and environmental feature sequence, and generating correlation features characterizing the mutual influence of multi-source time series data, includes: Based on a preset time resolution, the water condition feature sequence, engineering condition feature sequence, and environmental feature sequence are aligned along the time axis to construct a set of joint feature sequences with a unified time index; Multiple candidate time lag intervals are set within a preset time range. In each candidate time lag interval, the lag feature is constructed between two different feature elements in the joint feature sequence set, resulting in multiple sets of lag feature sequence pairs for multiple candidate time lag intervals. Based on each of the multiple sets of lagged feature sequence pairs, a consistency index is calculated within the corresponding time window; the consistency index is used to characterize the degree of correlation between feature sequences under different time lag conditions. Among multiple candidate time lag intervals, the candidate time lag interval with the optimal change consistency index is selected, and the selected candidate time lag interval is determined as the time lag relationship between the corresponding feature sequences. Based on the aforementioned time lag relationship and the corresponding consistency index of change, a correlation feature is constructed to characterize the mutual influence relationship between water condition characteristics, engineering condition characteristics and environmental characteristics.

4. The method according to claim 3, characterized in that, The change consistency index is obtained by weighting the trend consistency index, the magnitude correlation index, and the change rate matching index. Among them, the trend consistency index Satisfy the following formula: in, This represents the change in the characteristic sequence a at adjacent time points t. , This represents the eigenvalue of the characteristic sequence a at time point t. Let a represent the feature value of the feature sequence a at time point t-1. a and b are used to characterize any different class of water condition features, engineering condition features, or environmental features, respectively. This indicates that the lagged characteristic sequence b at time point The change at adjacent time points, , This indicates that the feature sequence b at time point eigenvalues, This indicates that the feature sequence b at time point eigenvalues, This refers to the time lag. This indicates an indicator function, which takes the value 1 when the condition in parentheses is true, and 0 otherwise; N is the number of valid time sample points participating in the calculation within the corresponding time window; The correlation index Satisfy the following formula: in, This represents the eigenvalue of the characteristic sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. denoted by Pearson correlation coefficient function, where a and b are used to characterize any different class of water condition characteristics, engineering condition characteristics, or environmental characteristics, respectively; The rate of change matching index Satisfy the following formula: in, This represents the eigenvalue of the characteristic sequence a at time point t. This indicates that the feature sequence b at time point eigenvalues, This is the time lag. This represents the rate of change of the characteristic sequence a at time point t. This indicates that the feature sequence b at time point The rate of change, This represents an exponential function.

5. The method according to claim 1, characterized in that, The process of reasoning through the multi-source feature sequences using a deep learning model to obtain reservoir operation status assessment results and trend prediction results includes: The multi-source feature sequences are input into the feature extraction branches corresponding to the convolutional neural network, recurrent neural network, and attention mechanism, respectively, for parallel processing, including: In the convolutional neural network branch, a one-dimensional convolution operation is used to perform sliding convolution processing on the multi-source feature sequence to extract the local change pattern of the feature sequence in different time windows and obtain the local temporal feature representation. In the recurrent neural network branch, the evolution of feature sequences over long time scales is modeled to obtain a global temporal feature representation; In the attention mechanism branch, importance weights for different time steps and different feature dimensions are calculated based on the multi-source feature sequence and its associated features, and the corresponding features are weighted to obtain the attention-enhanced feature representation; The local temporal feature representation, the global temporal feature representation, and the attention-enhanced feature representation are fused to obtain a fused feature representation. Based on the fused feature representation, the current operating status assessment results of the reservoir and the predicted trend of operating status changes within a preset time range are output through the fully connected layer.

6. The method according to claim 5, characterized in that, The step of fusing the local temporal feature representation, the global temporal feature representation, and the attention-enhanced feature representation to obtain a fused feature representation includes: Based on the local temporal feature representation, global temporal feature representation, and attention-enhanced feature representation, a feature importance evaluation matrix is ​​constructed; Based on the feature importance evaluation matrix, fusion weights corresponding to various feature representations are dynamically generated. Then, based on the fusion weights, the local temporal feature representation, the global temporal feature representation, and the attention-enhanced feature representation are weighted and fused to obtain the fused feature representation.

7. The method according to claim 6, characterized in that, The importance assessment matrix satisfies the following formula: in, Represents local temporal features; Represents global temporal features; This represents attention-enhancing feature representation; This is a trainable weight parameter matrix; Indicates feature concatenation operation; This represents element-wise multiplication and is used to characterize the interaction between different feature representations; This represents the normalization function.

8. The method according to claim 1, characterized in that, The process of determining the reservoir's operational status based on the reservoir's operational status assessment results and trend prediction results includes: The results of the reservoir operation status assessment are standardized to obtain the current status risk indicators. The trend prediction results are subjected to time decay weighting to obtain a trend risk index that characterizes the direction and intensity of changes in the reservoir's operating status. The current state risk index and the trend risk index are weighted and fused to obtain the comprehensive reservoir operation risk index; The comprehensive operational risk index of the reservoir is compared with a preset risk level threshold to determine the corresponding reservoir operational status level.

9. A reservoir operation status monitoring system based on deep learning, characterized in that, include: Data acquisition equipment and electronic equipment; The data acquisition equipment is used to collect multi-source time-series data on reservoir operation, including: hydrological time-series data, engineering time-series data, and environmental time-series data; wherein the hydrological time-series data includes: water level time-series data, rainfall time-series data, and seepage time-series data; the engineering time-series data includes: dam deformation time-series data and dam structural performance data; and the environmental time-series data includes: water quality data and meteorological data. The electronic device is used to construct a multi-source feature sequence of reservoir operation status based on the multi-source time-series data; to infer the multi-source feature sequence through a deep learning model to obtain reservoir operation status assessment results and trend prediction results; the deep learning model is composed of a convolutional neural network branch, a recurrent neural network branch, and an attention mechanism branch connected in parallel; and to determine the reservoir operation status based on the reservoir operation status assessment results and trend prediction results.

10. The system according to claim 9, characterized in that, The electronic device is also used for: The hydrological time series data is divided into time windows, and the water level change rate, cumulative rainfall and seepage change are extracted in each time window to form a hydrological feature sequence. Spatial grouping and temporal statistics are performed on the aforementioned engineering condition time series data to calculate the displacement changes and trends at each monitoring location of the dam, forming an engineering condition characteristic sequence; The environmental time series data are subjected to time statistical analysis to extract environmental state characteristics that reflect changes in the reservoir's operating environment, forming an environmental characteristic sequence; Based on the aforementioned hydrological characteristic sequence, engineering characteristic sequence, and environmental characteristic sequence, the temporal correlation and time lag relationship between different characteristic sequences are calculated to generate correlation features that characterize the mutual influence relationship of multi-source time series data. The hydrological feature sequence, engineering feature sequence, environmental feature sequence, and associated features are subjected to unified scale transformation and numerical normalization to obtain a multi-source feature sequence.

Images