Cascade power station water level estimation method and system based on deep learning

Abstract

The invention discloses a cascade power station water level estimation method and system based on deep learning, and belongs to the technical field of cascade power station water level estimation. Comprising the steps of creating a water level forecasting model set and forming upstream and downstream water level data sequences with complete time sequences; the upstream and downstream water level correlation intensity of different lag time periods is calculated, a correlation intensity sequence is generated, and the water level propagation time lag characteristic at the current moment and the rainfall response characteristic at the current moment are obtained; creating scene decision-making conditions, dividing a plurality of water level forecasting scenes, and determining the corresponding water level forecasting scene in combination with the water level propagation time-lag characteristics and the rainfall response characteristics; and matching an adaptive target water level forecasting model in the water level forecasting model set according to the determined water level forecasting scene. According to the method, the accuracy and stability of water level prediction in each scene are improved in a targeted manner, the adaptability of the model to complex operation working conditions is remarkably improved, and the problem that a traditional single model is insufficient in adaptability is solved.

Description

Technical Field

[0001] This invention relates to the technical field of cascade hydropower station water level estimation, specifically a deep learning-based method and system for cascade hydropower station water level estimation. Background Technology

[0002] Cascade hydropower stations, as an important component of integrated river basin development, typically undertake multiple key functions, including flood control, power generation, navigation, and water resource regulation. The operation and management of cascade hydropower stations directly affect the efficient utilization and safe scheduling of water resources in the basin. In particular, accurate water level prediction on ultra-short-term scales is one of the crucial foundational technologies for ensuring the safe and efficient operation of cascade hydropower stations.

[0003] Currently, the main technical methods for predicting water levels in cascade hydropower stations focus on using historical operational data to establish single-type water level forecasting models to predict the water level in front of the dam. However, these methods generally suffer from problems such as the simplification of prediction models and weak adaptability to different scenarios. In the actual operation of cascade hydropower stations, different hydrological conditions, meteorological conditions, and the operating status of the cascade hydropower stations interact and influence each other, resulting in complex water level change characteristics. For example, there is a significant time lag effect in the propagation of upstream water level fluctuations downstream, and the impact of rainfall across the basin on the water level in front of the dam also exhibits strong nonlinear characteristics. Furthermore, these water level influence characteristics also show significant differences under different seasons and time periods.

[0004] Existing traditional water level prediction methods generally fail to fully consider the differences in various scenarios during the operation of cascade hydropower stations, resulting in large prediction errors under complex conditions and making it difficult to meet the accuracy requirements of real-time operational decision-making for water level prediction. Furthermore, model selection lacks scientific decision-making basis, often relying solely on simple empirical rules or single data points, which cannot adapt to complex and ever-changing operational scenarios. Therefore, there is an urgent need to propose a technology with strong applicability that can adaptively select appropriate prediction models based on operational scenario conditions to improve the accuracy of ultra-short-term water level prediction and the level of operational management for cascade hydropower stations. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a method and system for estimating the water level of cascade hydropower stations based on deep learning.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for estimating the water level of cascade hydropower stations, comprising the following steps: A set of water level forecasting models is created based on the historical operation data and historical water level data of cascade hydropower stations. Continuous historical water level monitoring data of upstream and downstream stations are obtained at the current time and within the preset time period, forming a complete temporal sequence of upstream and downstream water level data. Based on upstream and downstream water level data sequences, the correlation strength between upstream and downstream water levels at different lag periods is calculated and a correlation strength sequence is generated. Based on the correlation strength sequence, the lag parameter that optimizes the correlation is determined, and the water level propagation time lag characteristics at the current moment are obtained. ; Based on historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data from representative water level stations for the corresponding time periods, the rainfall response characteristics at the current moment are obtained through correlation analysis. ; Scenario decision conditions are created, and several water level forecast scenarios are divided, taking into account the water level propagation time lag characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; Based on the determined water level forecast scenario, a suitable target water level forecast model is matched from the water level forecast model set; The current operating data of the cascade hydropower stations are input into the target prediction model, and the predicted water level in front of the cascade hydropower station dam is output.

[0007] In a further embodiment, the following steps are also included: Based on the dynamic updates of water level propagation time lag characteristics and rainfall response characteristics during the evolution stages of flood disasters, the forecast models are rematched and adapted according to each stage, and the corresponding upstream water level prediction results are output in stages.

[0008] In a further embodiment, the calculation process for the correlation intensity of upstream and downstream water levels at different lag periods is as follows: Based on the upstream and downstream water level data sequence, several candidate values ​​for the lag period to be evaluated are determined. For each lag period, candidate values The correlation strength between upstream and downstream water levels is calculated using the following formula. : ; In the formula, For a moment Water level data from upstream stations, For a moment Downstream sites after lag period The water level data after translation and These represent the average values ​​of the upstream and downstream water level sequences, respectively. The data length of the upstream and downstream water level data sequence; The correlation strength sequence is formed by calculating the correlation strength of the candidate values ​​for each lag period sequentially using the above formula. Correspondingly, the method for obtaining the water level propagation time delay characteristics at the current moment is as follows: Extreme value analysis was performed on the correlation strength sequence, and the candidate value of the lag period that maximizes the correlation strength was selected as the water level propagation time lag characteristic at the current moment. .

[0009] In a further embodiment, the characterization process of the rainfall response characteristics at the current moment is as follows: By aligning historical continuous rainfall monitoring data with monitoring data from representative water level stations for the corresponding time periods on a unified time scale, a rainfall-water level correspondence dataset is established. The rainfall-water level dataset is divided into several rainfall level intervals according to rainfall intensity. Determine the local response coefficient of rainfall to changes in water level in front of the dam. ; Based on the local response coefficients of each rainfall level range The comprehensive response characteristics are calculated using the following formula, which combines the weights of the interval data. : ;in, For the first The weight values ​​for each rainfall level interval. This represents the total number of rainfall level ranges. The comprehensive response characteristics Considered as the rainfall response characteristics at the current moment .

[0010] In a further embodiment, the scenario decision conditions include: water level propagation conditions and rainfall impact conditions; the creation steps are as follows: Based on historical data from continuous water level monitoring, the statistical average of the historical water level propagation time delay characteristic dataset is calculated. with standard deviation The water level propagation threshold is determined using the following formula. : In the formula, Water level propagation threshold Adjustment factor; Calculate the corresponding average value using historical data from continuous rainfall monitoring. with standard deviation The rainfall impact threshold is calculated using the following formula. : In the formula, Rainfall impact threshold Adjustment coefficient.

[0011] In a further embodiment, the water level forecast scenarios include at least: a scenario with insignificant water level propagation and weak rainfall impact, a scenario with significant water level propagation and weak rainfall impact, a scenario with insignificant water level propagation and significant rainfall impact, and a scenario with significant water level propagation and significant rainfall impact; the steps for classifying the water level forecast scenarios are as follows: like and This indicates that the current scenario is one where water level propagation is not significant and rainfall has a weak impact. like and This indicates that the current scenario is one of significant water level propagation and weak rainfall impact. like and This indicates that the current scenario is one where water level propagation is not significant but rainfall has a significant impact. like and This indicates that the current scenario is one where water level propagation is significant and rainfall has a significant impact. in, The water level propagation threshold, The threshold for the impact of rainfall.

[0012] In a further embodiment, the water level forecasting model set includes at least a basic water level forecasting model, an upstream water level coupled water level forecasting model, an inter-regional rainfall coupled water level forecasting model, and a two-factor coupled water level forecasting model; Among them, the basic water level forecast model is suitable for scenarios where water level propagation is not obvious and rainfall has a weak impact; the upstream water level coupled water level forecast model is suitable for scenarios where water level propagation is obvious and rainfall has a weak impact; the inter-regional rainfall coupled water level forecast model is suitable for scenarios where water level propagation is not obvious and rainfall has a significant impact; and the two-factor coupled water level forecast model is suitable for scenarios where water level propagation is obvious and rainfall has a significant impact.

[0013] In a further embodiment, the update process for the water level propagation time lag characteristics and rainfall response characteristics is as follows: Based on the disaster development rate at each stage of flood disaster evolution, the feature update cycle is dynamically set. ; in each update cycle Perform the following steps internally: The latest continuous monitoring data from upstream stations, downstream stations, and representative water level stations are obtained and added to the upstream and downstream water level data series. At the same time, historical data that exceeds the current time and the preset time period are removed to form a complete time-series updated upstream and downstream water level data series and an updated rainfall-water level dataset. The updated water level propagation time delay characteristics are obtained based on the updated upstream and downstream water level data sequences. , This refers to the number of updates; Determine updated rainfall response features based on the updated rainfall-water level dataset .

[0014] In a further embodiment, the following steps are also included: like and If the update is invalid, the current update will be considered invalid, and the previous update will be maintained. The corresponding features of the next step; otherwise, the update is deemed valid, triggering a new matching; where, For the first The time delay characteristics of water level propagation in this phase The threshold for the water level propagation time delay characteristic. For the first The characteristics of the rainfall response. This is the threshold for rainfall response characteristics.

[0015] A deep learning-based system for estimating the water level of cascade hydropower stations is provided to implement the water level estimation method for cascade hydropower stations described above, including: The first module is set to create a set of water level forecasting models based on the historical operating data and historical water level data of the cascade hydropower stations, obtain continuous historical water level monitoring data of upstream and downstream stations at the current time and within a preset time period, and form a complete time sequence of upstream and downstream water level data. The second module is configured to calculate the correlation strength between upstream and downstream water levels at different lag times based on upstream and downstream water level data sequences and generate a correlation strength sequence; based on the correlation strength sequence, it determines the lag parameter that optimizes the correlation and obtains the water level propagation time lag characteristics at the current moment. ; The third module is configured to calculate the rainfall response characteristics at the current moment by using historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data of representative water level stations for the corresponding time period through correlation analysis. ; The fourth module is set up to create scenario decision conditions and divides the scenario into several water level forecast scenarios, taking into account the water level propagation time delay characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; The fifth module is configured to match a suitable target water level forecasting model from the set of water level forecasting models based on the determined water level forecasting scenario. The sixth module is configured to input the current operating data of the cascade hydropower stations into the target prediction model and output the predicted water level in front of the cascade hydropower station dam. The seventh module is set to dynamically update the water level propagation time lag characteristics and rainfall response characteristics according to the evolution stage of flood disaster, re-match and adapt the forecast model according to the stage, and output the corresponding upstream water level prediction results in stages.

[0016] The beneficial effects of this invention are as follows: By establishing a basic water level forecasting model, an upstream water level coupled water level forecasting model, an inter-regional rainfall coupled water level forecasting model, and a two-factor coupled water level forecasting model, this invention forms a set of forecasting models that can comprehensively cover different scenarios of cascade hydropower station operation, specifically improve the accuracy and stability of water level prediction in each scenario, significantly improve the model's adaptability to complex operating conditions, and solve the problem of insufficient adaptability of traditional single models.

[0017] This invention extracts the water level propagation time delay characteristics and rainfall response characteristics at the current moment based on cross-correlation analysis and correlation analysis, respectively. Based on clear decision-making rules, it achieves accurate identification and quantitative judgment of the operation scenario of cascade hydropower stations, thereby ensuring the objectivity and accuracy of the prediction model selection and effectively reducing prediction errors caused by human experience interference.

[0018] This invention establishes water level propagation thresholds and rainfall impact thresholds based on historical data statistical analysis. By comparing quantitative features with these thresholds, scenarios are categorized, replacing the traditional method of selecting models based on manual experience and avoiding interference from subjective factors. Simultaneously, the feature update mechanism can dynamically adapt to changes in hydrological and meteorological conditions, ensuring the real-time nature and rationality of model selection and improving the intelligence level of cascade hydropower station scheduling. Attached Figure Description

[0019] Figure 1 This is a flowchart of the deep learning-based method for estimating the water level of a cascade hydropower station, as described in Example 1.

[0020] Figure 2 A comparison chart showing the maximum absolute error in the average water level prediction of the Majia River using different forecasting models.

[0021] Figure 3 A comparison chart showing the maximum absolute error of hourly water level predictions for the Majiahe River using different forecasting models.

[0022] Figure 4 A comparison chart of the average absolute error of the average water level prediction for Xintanba using different forecasting models.

[0023] Figure 5 A comparison chart showing the maximum absolute error of hourly water level predictions for Xintanba using different forecasting models. Detailed Implementation

[0024] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0025] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a deep learning-based method for estimating the water level of cascade hydropower stations, including the following steps: A set of water level forecasting models is created based on historical operating data and historical water level data of cascade hydropower stations. Continuous historical water level monitoring data of upstream and downstream stations at the current time and within a preset time period are obtained to form a complete temporal sequence of upstream and downstream water level data. It should be noted that the preset time period is the pre-set historical retrospective period.

[0026] Based on upstream and downstream water level data sequences, the correlation strength between upstream and downstream water levels at different lag periods is calculated and a correlation strength sequence is generated. Based on the correlation strength sequence, the lag parameter that optimizes the correlation is determined, and the water level propagation time lag characteristics at the current moment are obtained. ; Based on historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data from representative water level stations for the corresponding time periods, the rainfall response characteristics at the current moment are obtained through correlation analysis. ; Scenario decision conditions are created, and several water level forecast scenarios are divided, taking into account the water level propagation time lag characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; Based on the determined water level forecast scenario, a suitable target water level forecast model is matched from the water level forecast model set; The current operating data of the cascade hydropower stations are input into the target prediction model, and the predicted water level in front of the cascade hydropower station dam is output.

[0027] Taking the Xiluodu-Xiangjiaba cascade hydropower station as an example, the historical and current operating data mentioned in this embodiment are historically and currently collected operating data, respectively. Specifically, they include: inflow data, outflow data, and power output data of the cascade hydropower station. The corresponding water level data can be: water level data at upstream stations of the cascade hydropower station, water level data at downstream stations of the cascade hydropower station, rainfall within the cascade hydropower station area, and water level monitoring data in front of the dam. The collection frequency is generally set according to needs or specific scenarios, such as once every 15 minutes or once every hour. For example, the historical rainfall within the cascade hydropower station area is the historical rainfall data from typical rainfall stations set up within the cascade hydropower station area (including but not limited to Hengjiang Station and Pengshan-Gaochang Station), with a data frequency of 1 hour; the historical water level data in front of the dam is the real-time monitoring data from the Xiluodu dam-front automatic water level monitoring station, with a monitoring frequency of 15 minutes.

[0028] Traditional methods often fix a single or a small number of lag candidate values, which may miss the true time lag interval. This embodiment combines watershed characteristics to set a reasonable range and interval of candidate values, ensuring coverage of possible time lags in water flow propagation and avoiding misjudgments of time lags due to insufficient candidate value settings. The calculation process of the correlation strength between upstream and downstream water levels at different lag periods described in this embodiment is as follows: Continuous water level monitoring data from the upstream Majiahe hydrological station and the downstream Zhongxinchang station were obtained at the current time and within the previous 72 hours, with a uniform time resolution of 15 minutes, forming an upstream and downstream water level data sequence. ,in, ; Based on the characteristics of water flow propagation in the cascade hydropower station basin, several candidate values ​​for the lag period to be evaluated were determined. .like For each lag period, candidate values The correlation strength between upstream and downstream water levels is calculated using the following formula. : ; In the formula, For a moment Water level data from upstream stations, For a moment Downstream sites after lag period The water level data after translation and These represent the average values ​​of the upstream and downstream water level sequences, respectively. The data length of the upstream and downstream water level data sequence; furthermore, , The calculation method is the same and will not be elaborated further.

[0029] The correlation strength corresponding to the candidate values ​​for each lag period is calculated sequentially using the above formula, forming a correlation strength sequence, which is expressed as follows: .

[0030] The upstream water level changes in a cascade hydropower station exhibit a propagation delay to the downstream water level changes. Therefore, accurately determining the propagation time lag characteristics of the upstream and downstream water levels is crucial for the accuracy of water level estimation. This embodiment employs a cross-correlation analysis method to calculate the water level sequence of upstream stations within different lag periods. and downstream station water level sequence The correlation between them is used to determine the optimal time delay.

[0031] Correspondingly, the method for obtaining the water level propagation time delay characteristics at the current moment is as follows: Extreme value analysis (i.e., comparison of magnitude) is performed on the correlation strength sequence, and the candidate value of the lag period that maximizes the correlation strength is selected as the water level propagation lag characteristic at the current moment. .

[0032] In another embodiment, the characterization process of the rainfall response characteristics at the current moment is as follows: By aligning historical continuous rainfall monitoring data with monitoring data from representative water level stations for the corresponding time periods on a unified time scale, a rainfall-water level correspondence dataset is established. For example, linear interpolation is used to unify the frequency of water level monitoring data in front of the dam to 15 minutes, and regional historical rainfall data is also uniformly adjusted to a 15-minute frequency through interpolation to ensure the consistency of data processing scale.

[0033] The rainfall-water level dataset is divided into several rainfall level intervals according to rainfall intensity. Determine the local response coefficient of rainfall to changes in water level in front of the dam. This is used to indicate the specific degree of impact of rainfall on changes in the water level in front of the dam.

[0034] Furthermore, the local response coefficient described in this embodiment : In the formula, Indicates time Rainfall monitoring data, For a moment Water level monitoring data in front of the dam, , These represent the rainfall level ranges. Average rainfall and average water level within the area Rainfall level range The corresponding sample set; Based on the local response coefficients of each rainfall level range The comprehensive response characteristics are calculated using the following formula, which combines the weights of the interval data. : ;in, For the first The weight values ​​for each rainfall level interval. This represents the total number of rainfall level ranges. The comprehensive response characteristics Considered as the rainfall response characteristics at the current moment .

[0035] It should be noted that the first mentioned in this embodiment... Weight values ​​for each rainfall level range It is based on data subsets of each rainfall intensity level range. Number of data points within The calculation is as follows: ; In the formula, This represents the total number of data points across all rainfall intensity level ranges at the current moment.

[0036] By calculating local response coefficients in a hierarchical manner, the water level response patterns under different rainfall intensities are captured (e.g., the response coefficient for heavy rainfall is about 7 times that for light rainfall), avoiding feature distortion caused by averaging a single coefficient in traditional methods. The comprehensive response features integrate the influence weights of rainfall of different intensities, achieving a divide-and-conquer approach and overall aggregation representation. Compared with traditional methods, the matching degree between the features and the actual hydrological scenario is improved.

[0037] To address the issues of subjective and fixed threshold settings, limited decision-making dimensions, and insufficient robustness in traditional scenario decision-making, this embodiment proposes scenario decision conditions that include water level propagation conditions and rainfall impact conditions. The corresponding creation steps are as follows: Based on historical data from continuous water level monitoring, the statistical average of the historical water level propagation time delay characteristic dataset is calculated. with standard deviation The water level propagation threshold is determined using the following formula. : In the formula, Water level propagation threshold Adjustment factor; Calculate the corresponding average value using historical data from continuous rainfall monitoring. with standard deviation The rainfall impact threshold is calculated using the following formula. : In the formula, Rainfall impact threshold Adjustment coefficient.

[0038] It should be noted that the water level propagation threshold is... Adjustment coefficient and the rainfall impact threshold Adjustment coefficient Through multiple experiments, it was concluded that by monitoring the number of valid data points over three consecutive years of historical observation data, totaling 1000 data points, the value range can be determined to be 1.0 to 2.0, with 1.5 being the preferred value.

[0039] Based on continuous historical monitoring data, hydrological patterns are quantified through statistical averages and standard deviations. A threshold calculation model is then constructed using adjustment coefficients to replace traditional manual, experience-based value assignment. Threshold values ​​are derived entirely from data characteristics, avoiding inconsistencies in judgment standards caused by differences in the experience of different operators, thus ensuring uniformity in scenario-based decision-making.

[0040] Based on this, the water level forecast scenarios include at least: scenarios with insignificant water level propagation and weak rainfall impact, scenarios with significant water level propagation and weak rainfall impact, scenarios with insignificant water level propagation and significant rainfall impact, and scenarios with significant water level propagation and significant rainfall impact. Therefore, correspondingly, the water level forecast model set includes at least a basic water level forecast model, an upstream water level coupled water level forecast model, an inter-regional rainfall coupled water level forecast model, and a two-factor coupled water level forecast model. The steps for dividing the water level forecast scenario and matching the model are as follows: like and If the value is 0, it indicates that the current scenario is one where water level propagation is not significant and rainfall has a weak impact, and the basic water level forecast model should be used.

[0041] like and If the value is 0, it indicates that the current scenario is one of significant water level propagation and weak rainfall impact, and the upstream water level coupled with the water level forecasting model is used.

[0042] like and If , it indicates that the current scenario is one where water level propagation is not significant but rainfall has a significant impact; therefore, an interval rainfall coupled with a water level forecasting model is used.

[0043] like and This indicates that the current scenario is one of significant water level propagation and significant rainfall impact, and a two-factor coupled water level forecasting model is used, where... The water level propagation threshold, The threshold for the impact of rainfall.

[0044] To facilitate understanding of this embodiment, the basic water level forecasting model, the upstream water level coupled water level forecasting model, the interval rainfall coupled water level forecasting model, and the two-factor coupled water level forecasting model will be explained one by one below.

[0045] The basic water level forecasting model described in this embodiment is based on historical inflow, outflow, and power output data of cascade hydropower stations. It extracts the temporal correlation features between these three data points and changes in the water level upstream of the dam. A mapping relationship is constructed using a Long Short-Term Memory (LSTM) network to quantify the influence weights of flow and power output on the water level upstream of the dam at different operational stages. Ultimately, it achieves accurate prediction of the water level upstream of the dam based solely on current flow and power output data, suitable for stable operation scenarios without significant upstream water level propagation effects or significant rainfall impacts. The input variables of the LSTM network are three-dimensional temporal features, namely inflow, outflow, and power output data; the output variable is the predicted water level upstream of the dam. The gating structure and memory units in the LSTM network capture the dynamic changes in historical operating data of the cascade hydropower stations, completing the temporal expression of the water level change characteristics upstream of the dam under the combined effect of inflow and power output.

[0046] The construction process of the upstream water level coupled water level prediction model described in this embodiment is as follows: Based on the time delay characteristics of water level propagation Extract water level change characteristics and combine them with historical inflow, historical outflow and historical power plant output data to form a composite input set; A water level propagation characteristic coupling layer is constructed in a long short-term memory network. By extracting temporal features, the dynamic influence of upstream water level changes on the water level in front of the dam is obtained. It should be noted that the water level propagation characteristic coupling layer described in this embodiment is used to extract the temporal propagation pattern of upstream and downstream water levels. The propagation pattern (time delay effect, intensity attenuation, dynamic correlation) between upstream and downstream water levels is dynamically captured and quantified by a gating unit, and then synergistically integrated with operational data such as inflow, outflow, and power plant output.

[0047] Historical water level data in front of the cascade hydropower station dams were used as the training target for the model. The gradient descent algorithm was used to optimize the network parameters, thus forming the upstream water level coupled water level prediction model.

[0048] The upstream water level coupled water level forecasting model solves the problem of water level prediction under the upstream water level propagation effect. By strengthening the capture of time-delay features through the coupling layer, the prediction accuracy is improved in scenarios with obvious water level propagation compared with the basic water level forecasting model, ensuring accurate adaptation of the model to scenarios with obvious water level propagation and weak rainfall impact.

[0049] When the water level propagation time lag characteristic is lower than the set threshold (i.e., the upstream water level has no significant impact on the downstream propagation), and the rainfall response characteristic is higher than the set threshold (i.e., the interval rainfall has a significant impact on the water level in front of the dam), the interval rainfall coupled water level forecasting model is activated, such as in the case of local rainfall in the basin and no significant fluctuations in upstream water flow.

[0050] Rainfall-related characteristics: Based on historical rainfall data and corresponding water level changes in front of the dam within the cascade hydropower station area, rainfall response characteristics to water level are extracted (quantifying the correlation between rainfall intensity and water level changes). Operational characteristics: Includes real-time inflow and outflow data of cascade power stations and power station output data, covering the main influencing factors of power station scheduling and operation.

[0051] Feature combination: The above rainfall response features are integrated with inflow, outflow and power plant output data to form the input variable set of the model, ensuring coverage of both rainfall-driven and dispatch regulation factors.

[0052] The model is built on a Long Short-Term Memory (LSTM) network and includes an interaction layer for rainfall and flow characteristics. The interaction layer is designed to capture the nonlinear response relationship between rainfall intensity and power plant operating status. By co-processing rainfall response characteristics and operating data characteristics, the influence weights of inflow, outflow and power plant output adjustment on water level changes under different rainfall intensities are quantified, thus achieving the organic fusion of the two types of characteristics. Training process: Using historical measured water levels in front of the dam as the training target, the connection weights between the interaction layer and the output layer are iteratively optimized to enable the model to accurately map the correlation between input features and water levels in front of the dam. The real-time rainfall response characteristics, inflow, outflow and power plant output data are input into the trained model, and the model outputs the corresponding water level prediction results in front of the dam, which are used to reflect the water level change trend under the influence of rainfall in the interval.

[0053] Furthermore, when the water level propagation time lag characteristics are higher than the set threshold (i.e., the upstream water level has a significant impact on the downstream propagation), and the rainfall response characteristics are higher than the set threshold (i.e., the inter-regional rainfall has a significant impact on the water level in front of the dam), the two-factor coupled water level forecasting model is activated, such as the case of upstream flood discharge during the flood season superimposed with heavy rainfall in the basin, and the linkage between upstream and downstream water levels accompanied by inter-regional rainfall replenishment.

[0054] Water level propagation characteristics: Based on the water level propagation time lag characteristics determined by upstream and downstream water level data sequences, combined with real-time water level data from upstream stations (time-series aligned according to time lag characteristics), the propagation impact of upstream water level fluctuations on downstream is quantified. Rainfall-related characteristics: Based on rainfall response characteristics extracted from rainfall monitoring data in the cascade hydropower station area, the impact of inter-regional rainfall on the water level in front of the dam is quantified; Operational characteristics: Includes real-time inflow and outflow data of cascade hydropower stations and power station output data, covering the regulatory effect of dispatching behavior on water level; Feature combination: The above-mentioned water level propagation-related features and rainfall response features are integrated with inflow, outflow and power plant output data to form a multi-dimensional set of input variables for water level propagation, rainfall impact and operation scheduling.

[0055] The model is built on a Long Short-Term Memory (LSTM) network, with a core addition of a two-factor interaction layer (interaction module between water level propagation features and rainfall features). The role of the two-factor interaction layer is to capture the nonlinear synergistic relationship between the water level propagation effect and the rainfall impact: by synergistically processing the water level propagation-related features and rainfall response features, and combining the adjustment effect of operational data, the comprehensive impact weight of various factors on water level changes under different water level propagation intensities and different rainfall intensities is quantified, thereby achieving the organic fusion of multiple features; Using historical measured water levels in front of the dam as the training target, the model can accurately map the correlation between multidimensional input features and water levels in front of the dam by iteratively optimizing the connection weights of the two-factor interaction layer, LSTM unit, and output layer.

[0056] The real-time water level propagation characteristics, rainfall response characteristics, and power station operation data are input into the trained model, and the model outputs the corresponding water level prediction results in front of the dam, which are used to reflect the water level change trend under the combined influence of water level propagation and rainfall.

[0057] To more clearly illustrate the specific technical effects of determining the adaptation model based on water level propagation characteristics and rainfall response characteristics in this invention, we will take the ultra-short-term water level prediction of the Xiluodu-Xiangjiaba cascade hydropower station as an example.

[0058] like Figure 2 As shown, it can be clearly observed that during the non-flood season and without water discharge, the water level changes are relatively stable, and the basic water level-flow-output relationship can be used to make effective and accurate predictions. Therefore, in scenarios where water level propagation and rainfall response are not significant, the basic water level prediction model can meet the accuracy requirements and has adaptability.

[0059] like Figure 3 As shown, when there is a significant upstream and downstream water level propagation effect, the water level propagation characteristics cannot be accurately captured by the basic relationship alone. However, the upstream water level coupled water level forecasting model is closer to the actual water level change process because it takes into account the time lag characteristics of upstream and downstream water level propagation. Therefore, in scenarios where water level propagation is significant but rainfall response is not significant, the upstream water level coupled water level forecasting model is more adaptable.

[0060] according to Figure 4 As shown, under the scenario of significant rainfall impact, the overall forecast performance of the upstream water level coupled water level forecast model and the inter-regional rainfall coupled water level forecast model is similar. However, considering that the inter-regional rainfall coupled water level forecast model relies on less future data when the forecast performance is similar, the inter-regional rainfall coupled water level forecast model is more adaptable in the scenario where water level propagation is not obvious but rainfall impact is significant.

[0061] according to Figure 5 As shown, the two-factor coupled water level forecasting model has better adaptability under scenarios where water level propagation is significant and rainfall has a significant impact.

[0062] It should be noted that, Figures 2 to 5 In this embodiment, forecast model 1 refers to the basic water level forecast model, forecast model 2 refers to the upstream water level coupled water level forecast model, forecast model 3 refers to the interval rainfall coupled water level forecast model, and forecast model 4 refers to the two-factor coupled water level forecast model.

[0063] Through the above specific verification analysis, this invention clarifies the scenarios with different water level propagation characteristics and rainfall response characteristics, and rationally selects the basic water level forecasting model, upstream water level coupled water level forecasting model, interval rainfall coupled water level forecasting model, or two-factor coupled water level forecasting model that are suitable for them. It demonstrates the highly accurate matching relationship between each scenario and the corresponding model, which greatly improves the accuracy of ultra-short-term water level prediction for cascade power stations and ensures the refinement and reliability of power station operation and scheduling.

[0064] Example 2 While the method disclosed in Example 1 can select a model through scene matching, it does not consider the dynamic changes of key features during the evolution of flood disasters: the water propagation speed and the degree of rainfall impact will change significantly at different stages of flooding (such as the development period, peak period, and receding period). Specifically, during the development phase, the upstream water inflow increases rapidly, and the water level propagation lag may be shortened; during the peak period, rainfall is concentrated, and the rainfall response characteristics will be sharply enhanced; during the receding phase, the water flow slows down, the propagation lag is prolonged, and the impact of rainfall is weakened.

[0065] If the initial features and matching model are used, the prediction will be biased because the features are out of touch with the real-time conditions (for example, if the initial weak rainfall response feature matching model is still used during the peak period, the prediction accuracy will drop significantly).

[0066] Therefore, this embodiment also includes the following steps: Based on the dynamic updates of water level propagation time lag characteristics and rainfall response characteristics during the evolution stages of flood disasters, the forecast models are rematched and adapted according to each stage, and the corresponding upstream water level prediction results are output in stages.

[0067] The update process for water level propagation time lag characteristics and rainfall response characteristics is as follows: Based on the disaster development rate at each stage of flood disaster evolution, the feature update cycle is dynamically set. , Values ​​such as 30min, 1h, etc., ensure a balance between the timeliness and efficiency of feature updates.

[0068] In each update cycle Perform the following steps internally: The latest continuous monitoring data from upstream stations, downstream stations, and representative water level stations are obtained and added to the upstream and downstream water level data series. At the same time, historical data that exceeds the current time and the preset time period are removed (such as retaining the latest 72-hour data) to form a time-complete updated upstream and downstream water level data series and an updated rainfall-water level dataset. Based on the updated upstream and downstream water level data sequences, the corresponding steps in Example 1 are repeated to obtain the updated water level propagation time delay characteristics. , This refers to the number of updates; Based on the updated rainfall-water level dataset, repeat the corresponding steps in Example 1 to determine the updated rainfall response characteristics. .

[0069] In addition, to avoid invalid adaptations, the following steps are also included: like and If the update is invalid, the current update will be considered invalid, and the previous update will be maintained. The corresponding features of the next step; otherwise, the update is deemed valid, triggering a new matching; where, For the first The time delay characteristics of water level propagation in this phase The threshold for the water level propagation time delay characteristic. For the first The characteristics of the rainfall response. This is the threshold for rainfall response characteristics.

[0070] This embodiment improves prediction accuracy by dynamically adapting disaster stages, features, and models. It also reduces the amount of data processed per round by 80% and filters out more than 60% of invalid operations through lightweight processing such as incremental updates, dynamic cycles, and invalid filtering. This significantly reduces the data processing intensity, makes it suitable for deployment on ordinary industrial computers, and enhances engineering practicality.

[0071] Example 3 This embodiment discloses a deep learning-based cascade hydropower station water level estimation system, used to implement the cascade hydropower station water level estimation methods of Embodiments 1 and 2, including: The first module is set to create a set of water level forecasting models based on the historical operating data and historical water level data of the cascade hydropower stations, obtain continuous historical water level monitoring data of upstream and downstream stations at the current time and within a preset time period, and form a complete time sequence of upstream and downstream water level data. The second module is configured to calculate the correlation strength between upstream and downstream water levels at different lag times based on upstream and downstream water level data sequences and generate a correlation strength sequence; based on the correlation strength sequence, it determines the lag parameter that optimizes the correlation and obtains the water level propagation time lag characteristics at the current moment. ; The third module is configured to calculate the rainfall response characteristics at the current moment by using historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data of representative water level stations for the corresponding time period through correlation analysis. ; The fourth module is set up to create scenario decision conditions and divides the scenario into several water level forecast scenarios, taking into account the water level propagation time delay characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; The fifth module is configured to match a suitable target water level forecasting model from the set of water level forecasting models based on the determined water level forecasting scenario. The sixth module is configured to input the current operating data of the cascade hydropower stations into the target prediction model and output the predicted water level in front of the cascade hydropower station dam. The seventh module is set to dynamically update the water level propagation time lag characteristics and rainfall response characteristics according to the evolution stage of flood disaster, re-match and adapt the forecast model according to the stage, and output the corresponding upstream water level prediction results in stages.

Claims

1. A deep learning-based method for estimating the water level of cascade hydropower stations, characterized in that, Includes the following steps: A set of water level forecasting models is created based on the historical operation data and historical water level data of cascade hydropower stations. Continuous historical water level monitoring data of upstream and downstream stations are obtained at the current time and within the preset time period, forming a complete temporal sequence of upstream and downstream water level data. Based on upstream and downstream water level data sequences, the correlation strength between upstream and downstream water levels at different lag periods is calculated and a correlation strength sequence is generated. Based on the correlation strength sequence, the lag parameter that optimizes the correlation is determined, and the water level propagation time lag characteristics at the current moment are obtained. ; Based on historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data from representative water level stations for the corresponding time periods, the rainfall response characteristics at the current moment are obtained through correlation analysis. ; Scenario decision conditions are created, and several water level forecast scenarios are divided, taking into account the water level propagation time lag characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; Based on the determined water level forecast scenario, a suitable target water level forecast model is matched from the water level forecast model set; The current operating data of the cascade hydropower stations are input into the target prediction model, and the predicted water level in front of the cascade hydropower station dam is output.

2. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 1, characterized in that, It also includes the following steps: Based on the dynamic updates of water level propagation time lag characteristics and rainfall response characteristics during the evolution stages of flood disasters, the forecast models are rematched and adapted according to each stage, and the corresponding upstream water level prediction results are output in stages.

3. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 1, characterized in that, The calculation process for the correlation intensity of upstream and downstream water levels at different lag periods is as follows: Based on the upstream and downstream water level data sequence, several candidate values ​​for the lag period to be evaluated are determined. ; For each lag period, candidate values The correlation strength between upstream and downstream water levels is calculated using the following formula. : ； In the formula, For a moment Water level data from upstream stations, For a moment Downstream sites after lag period The water level data after translation and These represent the average values ​​of the upstream and downstream water level sequences, respectively. The data length of the upstream and downstream water level data sequence; The correlation strength sequence is formed by calculating the correlation strength of the candidate values ​​for each lag period sequentially using the above formula. Correspondingly, the method for obtaining the water level propagation time delay characteristics at the current moment is as follows: Extreme value analysis was performed on the correlation strength sequence, and the candidate value of the lag period that maximizes the correlation strength was selected as the water level propagation time lag characteristic at the current moment. .

4. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 1, characterized in that, The characterization process of the rainfall response characteristics at the current moment is as follows: By aligning historical continuous rainfall monitoring data with monitoring data from representative water level stations for the corresponding time periods on a unified time scale, a rainfall-water level correspondence dataset is established. The rainfall-water level dataset is divided into several rainfall level intervals according to rainfall intensity. Determine the local response coefficient of rainfall to changes in water level in front of the dam. ; Based on the local response coefficients of each rainfall level range The comprehensive response characteristics are calculated using the following formula, which combines the weights of the interval data. : ;in, For the first The weight values ​​for each rainfall level interval. This represents the total number of rainfall level ranges. The comprehensive response characteristics Considered as the rainfall response characteristics at the current moment .

5. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 1, characterized in that, The scenario decision conditions include: water level propagation conditions and rainfall impact conditions; the creation steps are as follows: Based on historical data from continuous water level monitoring, the statistical average of the historical water level propagation time delay characteristic dataset is calculated. with standard deviation The water level propagation threshold is determined using the following formula. : In the formula, Water level propagation threshold Adjustment factor; Calculate the corresponding average value using historical data from continuous rainfall monitoring. with standard deviation The rainfall impact threshold is calculated using the following formula. : In the formula, Rainfall impact threshold Adjustment coefficient.

6. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 1, characterized in that, The water level forecast scenarios include at least the following: scenarios with insignificant water level propagation and weak rainfall impact, scenarios with significant water level propagation and weak rainfall impact, scenarios with insignificant water level propagation and significant rainfall impact, and scenarios with significant water level propagation and significant rainfall impact; the steps for classifying the water level forecast scenarios are as follows: like and This indicates that the current scenario is one where water level propagation is not significant and rainfall has a weak impact. like and This indicates that the current scenario is one of significant water level propagation and weak rainfall impact. like and This indicates that the current scenario is one where water level propagation is not significant but rainfall has a significant impact. like and This indicates that the current scenario is one where water level propagation is significant and rainfall has a significant impact. in, The water level propagation threshold, The threshold for the impact of rainfall.

7. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 6, characterized in that, The water level forecasting model set includes at least a basic water level forecasting model, an upstream water level coupled water level forecasting model, an inter-regional rainfall coupled water level forecasting model, and a two-factor coupled water level forecasting model; Among them, the basic water level forecast model is suitable for scenarios where water level propagation is not obvious and rainfall has a weak impact; the upstream water level coupled water level forecast model is suitable for scenarios where water level propagation is obvious and rainfall has a weak impact; the inter-regional rainfall coupled water level forecast model is suitable for scenarios where water level propagation is not obvious and rainfall has a significant impact; and the two-factor coupled water level forecast model is suitable for scenarios where water level propagation is obvious and rainfall has a significant impact.

8. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 2, characterized in that, The update process for water level propagation time lag characteristics and rainfall response characteristics is as follows: Based on the disaster development rate at each stage of flood disaster evolution, the feature update cycle is dynamically set. ; in each update cycle Perform the following steps internally: The latest continuous monitoring data from upstream stations, downstream stations, and representative water level stations are obtained and added to the upstream and downstream water level data series. At the same time, historical data that exceeds the current time and the preset time period are removed to form a complete time-series updated upstream and downstream water level data series and an updated rainfall-water level dataset. The updated water level propagation time delay characteristics are obtained based on the updated upstream and downstream water level data sequences. , This refers to the number of updates; Determine updated rainfall response features based on the updated rainfall-water level dataset .

9. The deep learning-based method for estimating the water level of cascade hydropower stations according to claim 7, characterized in that, It also includes the following steps: like and If the update is invalid, the current update will be considered invalid, and the previous update will be maintained. Corresponding features of the next step; Conversely, if the update is not valid, a new matching process is initiated; where, For the first The time delay characteristics of water level propagation in this phase The threshold for the water level propagation time delay characteristic. For the first The characteristics of the rainfall response. This is the threshold for rainfall response characteristics.

10. A deep learning-based cascade hydropower station water level estimation system, used to implement the cascade hydropower station water level estimation method as described in any one of claims 1 to 9, characterized in that, include: The first module is set to create a set of water level forecasting models based on the historical operating data and historical water level data of the cascade hydropower stations, obtain continuous historical water level monitoring data of upstream and downstream stations at the current time and within a preset time period, and form a complete time sequence of upstream and downstream water level data. The second module is configured to calculate the correlation strength between upstream and downstream water levels at different lag times based on upstream and downstream water level data sequences and generate a correlation strength sequence; based on the correlation strength sequence, it determines the lag parameter that optimizes the correlation and obtains the water level propagation time lag characteristics at the current moment. ; The third module is configured to calculate the rainfall response characteristics at the current moment by using historical continuous rainfall monitoring data of the cascade hydropower station area and monitoring data of representative water level stations for the corresponding time period through correlation analysis. ; The fourth module is set up to create scenario decision conditions and divides the scenario into several water level forecast scenarios, taking into account the water level propagation time delay characteristics. and rainfall response characteristics Determine the corresponding water level forecast scenario; The fifth module is configured to match a suitable target water level forecasting model from the set of water level forecasting models based on the determined water level forecasting scenario. The sixth module is configured to input the current operating data of the cascade hydropower stations into the target prediction model and output the predicted water level in front of the cascade hydropower station dam. The seventh module is set to dynamically update the water level propagation time lag characteristics and rainfall response characteristics according to the evolution stage of flood disaster, re-match and adapt the forecast model according to the stage, and output the corresponding upstream water level prediction results in stages.

Images