Hydropower station water level detection method and system

Abstract

The invention discloses a hydropower station water level detection method and system. The method comprises the steps that an original water flow characteristic parameter set is obtained through multiple sets of Doppler effect radar flow measuring devices; performing segmentation extraction to obtain a water flow feature sub-data set; inputting a Transform network coding layer to generate a water flow feature coding vector; outputting a preliminary water level predicted value sequence through a decoding layer; a radar flow measurement model is called for correction to obtain a correction sequence; time sequence integration is carried out to obtain a water level change curve. The system comprises a multi-source Doppler radar signal acquisition unit, a spatial-temporal feature segmentation extraction unit, a Transform coding processing unit, a water level preliminary prediction unit, a radar model correction unit and a time sequence integration output unit. According to the method and system, multi-source parameters are deeply fused, the environmental adaptability is improved, the continuity and reliability of a detection result are enhanced, the water level of the hydropower station can be accurately monitored, and support is provided for scheduling decision and safety prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of water level detection of hydropower stations, and in particular to a water level detection method and system for hydropower stations. BACKGROUND

[0002] In the operation process of a hydropower station, water level detection is a key link for ensuring the safe and stable operation of the power station, optimizing water resource scheduling, and achieving efficient power generation. With the rapid development of the hydropower industry, higher requirements are placed on the real-time, accuracy, and reliability of water level detection. Traditional water level detection methods rely mainly on a single sensor or a simple model, which is difficult to cope with the challenges brought by the dynamic changes of multiple parameters such as water flow speed, direction, and disturbance under complex hydrological conditions. Especially in the flood season or dry season and other extreme hydrological conditions, the water level fluctuates dramatically, and the existing detection methods often fail to fully capture the dynamic characteristics of the water body, resulting in detection results that are lagging or biased, which affects the scheduling decisions and safety control of the power station.

[0003] The existing technology has two significant shortcomings: on the one hand, the detection model lacks the ability to integrate multiple hydrological parameters, and mostly uses single sensor data or simple parameter combinations for water level calculation, failing to fully exploit the correlation between parameters such as water flow speed, direction, and disturbance frequency, resulting in a decline in detection accuracy when the water flow state is complex; on the other hand, the adaptability is poor, and it is difficult to dynamically adjust the detection strategy according to the specific environmental characteristics of the hydropower station (such as dam structure, basin precipitation, etc.), and when facing environmental disturbances such as water temperature changes and sediment content differences, the model correction mechanism is imperfect, and the detection results are easily affected by external factors, which cannot stably meet the water level monitoring needs under different working conditions. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a water level detection method and system for a hydropower station.

[0005] In a first aspect, a water level detection method for a hydropower station is provided, comprising the following steps: Step S1: continuously capturing the dynamic water flow signals formed in the water flow process in the monitoring area by multiple Doppler effect radar flow measurement devices, and obtaining a set of original water flow characteristic parameters including water flow speed, water flow direction, and water flow disturbance frequency; Step S2: based on a preset time-space sampling interval, segmenting and extracting the obtained set of original water flow characteristic parameters to obtain a plurality of water flow characteristic sub-data sets with time series attributes, each sub-data set corresponding to a continuous time-space sampling unit, and each sub-data set including the dynamic change trajectory of the water flow characteristic parameters in the sampling unit; Step S3, input the obtained water flow feature sub-data set into the encoding layer of the water level detection model based on the pre-constructed Transformer network framework, mine the correlation between different parameters in the water flow feature sub-data set through the multi-head self-attention mechanism in the encoding layer, and generate a water flow feature encoding vector including parameter correlation weights; Step S4, input the generated water flow feature encoding vector into the decoding layer of the Transformer network framework, combine the preset water level reference parameters of the hydropower station, and perform layer-by-layer analysis on the encoding vector to output a preliminary water level prediction value sequence; Step S5, call the radar flow measurement model based on the Doppler effect to correct the output preliminary water level prediction value sequence, compare and analyze the preliminary prediction value with the water flow velocity parameter in the original water flow feature parameter set in step S1 through the preset water flow-water level coupling relationship in the model, generate a water level correction coefficient, and modify the preliminary water level prediction value sequence according to the coefficient; Step S6, time sequence integration is performed on the modified water level prediction value sequence to obtain a water level change curve of the hydropower station in the monitoring area in a continuous time period. Each data point in the curve includes a corresponding time stamp and a water level value.

[0006] Further, in step S2, when the original water flow feature parameter set is segmented and extracted, the precipitation parameters of the river basin where the hydropower station is located are also dynamically adjusted, the precipitation parameters including precipitation intensity and precipitation duration, and the spatiotemporal sampling interval is determined by the following model formula: ; Wherein, is the adjusted spatiotemporal sampling interval, is the initial spatiotemporal sampling interval, is a precipitation influence factor, with a value range of 0.005-0.02, is the precipitation intensity, i.e., the amount of precipitation per unit time, is the precipitation duration.

[0007] Further, in step S3, when the encoding layer generates the water flow feature encoding vector, historical monitoring parameters of the water level of the hydropower station are also introduced, the historical monitoring parameters including water level peak value, water level valley value and water level change rate in a past preset time period, and the output result of the multi-head self-attention mechanism is optimized by the following model formula: ; Wherein, is the optimized water flow feature encoding vector, is the initial encoding vector output by the multi-head self-attention mechanism, is a historical parameter influence factor, with a value range of 0.1-0.3, is a historical same period water level value corresponding to the current time, is an average water level value in the past preset time period, is the highest water level value in the past preset time period, is the lowest water level value in the past preset time period.

[0008] Further, in step S4, the decoding layer of the Transformer network framework combines the dam structure parameters of the hydropower station when outputting the preliminary water level prediction value sequence, the dam structure parameters including dam height, dam water-facing slope and dam drainage outlet distribution density, and calculates the preliminary water level prediction value through the following model formula: ; wherein, is the preliminary water level prediction value, is the analysis function of the decoding layer, is the water flow feature encoding vector, is the structure parameter influence coefficient, the value range being 0.05-0.2, is the dam water-facing slope value, is the dam drainage outlet distribution density, i.e. the number of drainage outlets per unit area, is the dam height.

[0009] Further, in step S5, when the radar flow measurement model generates the water level correction coefficient, the water temperature parameter and the water sediment content parameter are introduced, and the water level correction coefficient is calculated through the following model formula: ; wherein, is the water level correction coefficient, is the temperature influence coefficient, the value range being 0.01-0.03, is the current water temperature, is the standard reference water temperature, is the sediment content influence coefficient, the value range being 0.02-0.05, is the water sediment content, i.e. the mass of sediment per unit volume of water.

[0010] Further, in step S6, when the time series integration is performed on the corrected water level prediction value sequence, the hydropower generator set operation parameters are introduced, the operation parameters including generator set power generation power and generator set operation number, and the integrated water level change curve is secondarily optimized through the following model formula: ; wherein, is the secondary optimized water level value, is the time series integrated water level value, This is the unit impact factor, with a value ranging from 0.01 to 0.03. This refers to the power generation capacity of a single generating unit. Number of operating units This represents the maximum total power generation capacity of the hydropower station units.

[0011] Further, step S3 includes the following sub-steps: S31, aligning the parameters of the water flow feature subset obtained in step S2, transforming different types of water flow feature parameters to the same data dimension space, so that the parameters of water flow velocity, water flow direction, and water flow disturbance frequency have a unified dimension. The dimension transformation process determines the transformation coefficients based on the distribution range of each parameter in historical data; S32, inputting the dimension-aligned water flow feature subset into the embedding layer of the Transformer network framework's encoding layer in chronological order. The embedding layer converts each parameter value into a high-dimensional vector representation, with each parameter corresponding to a vector dimension... The degree is determined based on the dynamic range of parameter changes; the larger the range of changes, the higher the corresponding vector dimension. In S33, the high-dimensional vector output from the embedding layer is input into the multi-head self-attention mechanism. This mechanism calculates the attention weights between different parameter vectors through multiple parallel attention heads. Each attention head focuses on different types of correlations between parameters, including linear and non-linear correlations. In S34, the multiple attention weight matrices output by the multi-head self-attention mechanism are weighted and summed to obtain a comprehensive attention weight matrix. This matrix is ​​then multiplied with the high-dimensional vector output from the embedding layer to generate a water flow feature encoding vector that includes parameter correlation weights.

[0012] Further, step S4 includes the following sub-steps: S41, concatenating the water flow feature encoding vector generated in step S3 with the preset hydropower station water level reference parameter vector to form the input vector of the decoding layer, wherein the water level reference parameter vector includes vector representations of the hydropower station's normal storage water level, dead water level, and design flood level parameters; S42, inputting the concatenated input vector into the first feedforward neural network layer of the Transformer network framework decoding layer, which performs feature mapping on the input vector through linear transformation and nonlinear activation functions to enhance the feature mapping of the vector with respect to water flow. S43, input the output of the first feedforward neural network layer to the residual connection and layer normalization module of the decoding layer. The original features of the input vector are preserved through the residual connection, and the distribution of the feature vector is adjusted through the layer normalization so that the mean and variance of each element in the vector are within the preset range. S44, input the feature vector after the residual connection and layer normalization process to the second feedforward neural network layer of the decoding layer. The second feedforward neural network layer further extracts features from the feature vector through multiple sets of convolution kernels and outputs a preliminary water level prediction value sequence.

[0013] Further, step S5 includes the following sub-steps: S51, extracting the dynamic change sequence of water flow velocity parameters from the original water flow characteristic parameter set obtained in step S1, and synchronizing this sequence with the preliminary water level prediction value sequence output in step S4 to ensure that the data points in the two sequences correspond one-to-one in the time dimension; S52, calling the water flow-water level coupling relationship module in the radar flow measurement model based on the Doppler effect, which establishes a correlation model between water flow velocity and water level according to a preset physical equation, inputting the synchronized water flow velocity parameters into the correlation model to obtain the corresponding theoretical water level value sequence; 53. Calculate the deviation between the preliminary water level prediction value sequence output in step S4 and the theoretical water level value sequence obtained in step S52. Arrange the deviation values ​​in chronological order to form a deviation sequence, and smooth the deviation sequence using a sliding window algorithm to obtain the smoothed deviation value. S54. Calculate the water level correction coefficient based on the smoothed deviation value. Specifically, the smoothed deviation value at each time point is compared with the theoretical water level value at that time point to obtain the correction coefficient for that time point. The correction coefficients at all time points are combined into a water level correction coefficient sequence, which is used to correct the preliminary water level prediction value sequence.

[0014] Secondly, a hydropower station water level detection system is provided, comprising: A multi-source Doppler radar signal acquisition unit, which is connected to multiple Doppler effect radar flow measurement devices deployed in the predetermined monitoring area of ​​the hydropower station, is used to receive and store the original water flow characteristic parameter set captured by the radar flow measurement devices; The spatiotemporal feature segmentation extraction unit is connected to the multi-source Doppler radar signal acquisition unit. It is used to segment the original water flow feature parameter set according to the preset spatiotemporal sampling interval to generate multiple water flow feature subsets. The Transformer encoding processing unit, connected to the spatiotemporal feature segmentation extraction unit, is used to mine the correlation between water flow feature subsets through a multi-head self-attention mechanism and output water flow feature encoding vectors. The preliminary water level prediction unit is connected to the Transformer encoding processing unit. It is used to analyze the water flow feature encoding vector in combination with the hydropower station water level reference parameters and output a preliminary water level prediction value sequence. The radar model correction unit is connected to the preliminary water level prediction unit and the multi-source Doppler radar signal acquisition unit, respectively. It is used to call the radar flow measurement model based on the Doppler effect to correct the preliminary water level prediction value sequence and generate the corrected water level prediction value sequence. The time-series integration output unit is connected to the radar model correction unit. It is used to integrate the corrected water level prediction value sequence in time and output the water level change curve of the hydropower station. This unit is also connected to the monitoring terminal of the hydropower station to transmit the water level change curve to the monitoring terminal for display.

[0015] The beneficial effects of the technical solution of this invention: This paper presents a method and system for water level detection in hydropower stations. To improve detection accuracy, multiple Doppler effect radar flow measurement devices capture multi-dimensional raw parameters such as water flow velocity, direction, and disturbance frequency. These parameters are then extracted into feature subsets through spatiotemporal segmentation. Furthermore, the multi-head self-attention mechanism of the encoding layer in the Transformer network framework is used to deeply mine the correlations between parameters and generate encoding vectors. This solves the problem of insufficient fusion of multi-source parameters in existing technologies, allowing full utilization of the potential connections between parameters and laying the foundation for accurate prediction. In terms of enhancing adaptability, the decoding layer combines water level benchmark parameter analysis with parameters such as historical water levels and dam structure to optimize prediction. During radar flow measurement model calibration, parameters such as water temperature and sediment content are incorporated to generate correction coefficients. The sampling interval can also be dynamically adjusted based on precipitation parameters. Combined with optimized integration of unit operating parameters, the system improves the mechanism for coping with complex environments and overcomes the poor adaptability of existing technologies. The continuous water level change curve obtained through time-series integration, with data points including timestamps and water level values, can stably cope with various interferences, providing accurate and reliable data for power station scheduling decisions and safety control, and comprehensively meeting the water level monitoring needs under different operating conditions. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of the hydropower station water level detection method of the present invention; Figure 2 This is a structural diagram of the hydropower station water level detection system of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a method for detecting water level in a hydropower station includes the following steps: Step S1: By deploying multiple sets of Doppler effect radar flow measurement devices in the predetermined monitoring area of ​​the hydropower station, the dynamic water flow signals formed during the water flow process in the monitoring area are continuously captured, and the original water flow characteristic parameter set including water flow velocity, water flow direction and water flow disturbance frequency is obtained. The signal capture direction of each set of radar flow measurement devices forms a preset angle with the water surface, and the capture range of each set of devices has a preset overlapping area. Specifically, step S1 is the foundational data acquisition stage of the entire water level detection process. Its core lies in acquiring comprehensive and accurate raw water flow characteristic parameters through multiple sets of Doppler effect radar flow measurement devices. The technical parameters of this step involve the number of radar flow measurement devices deployed, the angle between the signal acquisition direction and the water surface, and the overlapping area of ​​the acquisition range. Deploying multiple devices allows for the acquisition of water flow signals from different angles and positions, reducing potential blind spots for single devices and improving data integrity. The preset angle between the signal acquisition direction and the water surface is typically between 30 and 60 degrees. This angle range ensures that the radar signal effectively penetrates the disturbance layer on the water surface while accurately capturing the dynamic characteristics of the water flow. The preset overlapping area of ​​the acquisition range is generally set to 10% to 20% of the device's monitoring range. Data comparison and verification within the overlapping area improves the reliability of the raw data, providing a solid foundation for subsequent data analysis and processing. The significance of this step lies in providing first-hand raw data reflecting the true flow state of the water body for the entire water level detection method. This data serves as the basis for all subsequent analyses and predictions.

[0019] In the specific implementation process, the number of Doppler effect radar flow measurement devices to be deployed is first determined based on the size and terrain features of the monitoring area of ​​the hydropower station. Generally, for monitoring areas of small and medium-sized hydropower stations, deploying 3 to 5 sets of devices is sufficient; for large hydropower stations or monitoring areas with complex terrain, 5 to 8 sets or even more devices are required. Each set of devices is installed on fixed supports around the monitoring area. The height of the supports is adjusted according to the monitoring range and water depth, usually between 5 and 15 meters. The signal acquisition direction of each set of devices is adjusted so that it forms an angle of about 45 degrees with the water surface. This angle has been verified through multiple practices and can maximize the acquisition of water flow speed and direction information while ensuring signal strength. The acquisition range of each set of devices is set to ensure that there is a 15% overlap area between the acquisition ranges of adjacent devices. For example, if the monitoring radius of a set of devices is 100 meters, then the radius of its overlap area with adjacent devices is 15 meters. After the device is started, it captures water flow signals at a frequency of 10 times per second, continuously acquiring raw data including water flow velocity, water flow direction, and water flow disturbance frequency. The monitoring range for water flow velocity is 0.1 m / s to 10 m / s, the monitoring accuracy for water flow direction is ±5 degrees, and the monitoring range for water flow disturbance frequency is 0.1 Hz to 10 Hz. This raw data is transmitted in real time to the data storage unit, where it is classified and stored in chronological order to prepare for subsequent processing steps.

[0020] Step S2: Based on the preset spatiotemporal sampling interval, the original water flow feature parameter set obtained in step S1 is segmented and extracted to obtain multiple water flow feature subsets with time series attributes. Each subset corresponds to a continuous spatiotemporal sampling unit, and each subset includes the dynamic change trajectory of the water flow feature parameters within the sampling unit. Specifically, step S2 is a crucial step in processing the original set of water flow feature parameters. Its main purpose is to segment and extract the continuous raw data according to a preset spatiotemporal sampling interval, obtaining a water flow feature subset with time-series attributes. The technical parameters of this step include the size of the spatiotemporal sampling interval, which directly affects the accuracy and efficiency of subsequent data analysis. If the spatiotemporal sampling interval is too short, the data volume will be too large, increasing the computational burden of subsequent processing; if the interval is too long, important water flow feature information may be lost, affecting the accuracy of water level prediction. Therefore, it is necessary to reasonably set the spatiotemporal sampling interval based on the dynamic changes in the hydropower station's water flow. The significance of this step lies in orderly dividing the massive and continuous raw data, transforming it into subsets that are easy to analyze and process. Each subset corresponds to a specific spatiotemporal unit, reflecting the dynamic trajectory of water flow characteristics within that unit, laying the foundation for subsequent input into the Transformer network framework for processing.

[0021] In the specific implementation process, the initial spatiotemporal sampling interval is first determined based on the analysis results of historical water flow data of the hydropower station. For areas with relatively gentle water flow changes, the spatiotemporal sampling interval can be set to 10 minutes × 100 square meters; for areas with drastic water flow changes, such as near the dam or the intake area, the spatiotemporal sampling interval is set to 5 minutes × 50 square meters. Then, based on the set spatiotemporal sampling interval, the original water flow feature parameter set obtained in step S1 is segmented and extracted. Taking a spatiotemporal sampling interval of 5 minutes × 50 square meters as an example, the continuous original data is divided into a time unit of 5 minutes and a spatial unit of 50 square meters. The original data in each spatiotemporal unit constitutes a water flow feature subset. During the extraction process, it is necessary to ensure that each subset contains the complete dynamic trajectory of parameters such as water flow velocity, water flow direction, and water flow disturbance frequency within that spatiotemporal unit. For example, within a 5-minute time unit, it should include 300 data points (5 minutes × 60 seconds / minute × 1 occurrence / second) of water flow velocity, 300 data points of water flow direction, and 300 data points of water flow disturbance frequency. Simultaneously, a corresponding spatiotemporal identifier is added to each subset, including a timestamp and spatial location information. The timestamp is accurate to the second, and the spatial location information is represented in coordinate form, ensuring that the spatiotemporal attributes of the subset are clear and unambiguous. After extraction, an integrity check is performed on each subset. If any data is missing or abnormal, it is promptly marked for targeted processing in subsequent steps.

[0022] Step S3: Input the water flow feature subset obtained in step S2 into the encoding layer of the pre-built water level detection model based on the Transformer network framework. The multi-head self-attention mechanism in the encoding layer is used to mine the correlation between different parameters in the water flow feature subset and generate a water flow feature encoding vector including parameter correlation weights. The number of attention heads in the multi-head self-attention mechanism is adaptively adjusted according to the complexity of the historical water flow feature data of the hydropower station. Specifically, step S3 is the core step in deep processing of the water flow feature subset using the Transformer network framework. Its main task is to mine the correlation between water flow feature parameters through a multi-head self-attention mechanism within the encoding layer, generating water flow feature encoding vectors that include parameter correlation weights. The technical parameters of this step include the number of attention heads in the multi-head self-attention mechanism, which is adaptively adjusted based on the complexity of the historical water flow feature data from the hydropower station. When the historical data is highly complex, i.e., the water flow feature parameters change frequently and have complex correlations, a larger number of attention heads is needed to fully capture the subtle correlations between different parameters; conversely, the number of attention heads can be appropriately reduced to decrease computational complexity. The significance of this step lies in transforming the original parameter information in the water flow feature subset into encoding vectors with deep semantics and correlations. These vectors can more effectively represent the intrinsic relationship between water flow features and water level, providing high-quality feature input for subsequent water level prediction.

[0023] In the specific implementation process, the water flow feature subsets obtained in step S2 are first input into the encoding layer of the water level detection model based on the Transformer network framework in chronological order. The input data dimension of the encoding layer is determined according to the number of parameters and the amount of data in the subset. For example, if each water flow feature subset contains three parameters—water flow velocity, water flow direction, and water flow disturbance frequency—and each parameter has 300 data points, then the input dimension is 3×300. Then, the number of attention heads in the multi-head self-attention mechanism is determined according to the complexity of the historical water flow feature data of the hydropower station. If the water flow parameters in the historical data are relatively complex, such as the water flow data during the flood season, the number of attention heads is set to 16; if the water flow parameters in the historical data are relatively stable, such as the water flow data during the dry season, the number of attention heads is set to 8. The multi-head self-attention mechanism calculates the attention weights between different parameter vectors through multiple parallel attention heads. Each attention head focuses on different types of parameter relationships. For example, some attention heads focus on the relationship between water flow velocity and water flow direction, while others focus on the relationship between water flow velocity and water flow disturbance frequency. During the calculation, the hidden layer dimension of each attention head is allocated according to the input dimension and the number of attention heads. For example, if the input dimension is 900 (3×300) and the number of attention heads is 16, then the hidden layer dimension of each attention head is 56 (900÷16≈56). Finally, the attention weight matrices output by multiple attention heads are weighted and summed to obtain a comprehensive attention weight matrix. This matrix is ​​then multiplied with the input water flow feature subset vector to generate a water flow feature encoding vector of dimension 512. This vector contains the correlation weight information between various water flow parameters.

[0024] Step S4: Input the water flow feature encoding vector generated in step S3 into the decoding layer of the Transformer network framework. The decoding layer combines the preset hydropower station water level benchmark parameters to parse the encoding vector layer by layer and output a preliminary water level prediction value sequence. Each prediction value in the sequence corresponds to the time node of each water flow feature subset in step S2. Specifically, step S4 is a crucial step in the Transformer network framework for water level prediction. Its function is to input the water flow feature encoding vector generated by the encoding layer into the decoding layer, and then perform layer-by-layer analysis based on preset hydropower station water level benchmark parameters to output a preliminary sequence of predicted water levels. The technical parameters for this step include the selection and setting of water level benchmark parameters. These parameters typically include the normal storage level, dead storage level, and design flood level of the hydropower station, which are important references for the decoding layer's analysis. The normal storage level is the water level maintained by the hydropower station during normal operation to ensure power generation and water supply; the dead storage level is the lowest water level allowed to recede under normal operating conditions; and the design flood level is the highest water level reached in front of the dam when the reservoir encounters a design flood. The accurate setting of these benchmark parameters directly affects the accuracy of the decoding layer's analysis and the rationality of the preliminary water level predictions. The significance of this step lies in transforming the water flow feature information contained in the encoding vector into specific predicted water levels, providing preliminary prediction results for subsequent correction steps.

[0025] In the specific implementation process, the preset water level reference parameters are first retrieved from the hydropower station's database, including the normal storage water level of 150 meters, the dead water level of 120 meters, and the design flood level of 160 meters. These parameters are then converted into a vector representation with the same dimension as the water flow feature encoding vector, for example, a water level reference parameter vector with a dimension of 512. Next, the water flow feature encoding vector generated in step S3 is concatenated with the water level reference parameter vector to form a decoding layer input vector with a dimension of 1024 (512+512). This input vector is then fed into the first feedforward neural network layer of the decoding layer, which contains 2048 neurons. A linear transformation maps the input vector from 1024 dimensions to 2048 dimensions, followed by processing using the ReLU nonlinear activation function to enhance the feature signals in the vector related to water level prediction. The processed vector is input to the residual connection and layer normalization module. The residual connection adds the input vector to the output vector of the feedforward neural network layer, preserving the original features of the input vector. Layer normalization standardizes the vector, making the mean of each element in the vector 0 and the variance 1, ensuring the stability of the feature distribution. Next, the processed vector is input to the second feedforward neural network layer of the decoding layer. This layer contains 1024 neurons, which uses a linear transformation to map the vector from 2048 dimensions to 1 dimension, outputting the preliminary water level prediction value corresponding to each time node. Finally, a sequence of preliminary water level prediction values ​​corresponding to the time nodes of the water flow feature subset in step S2 is obtained. The range of each prediction value in the sequence is between the dead water level of 120 meters and the design flood level of 160 meters, with the prediction accuracy controlled within ±0.5 meters.

[0026] Step S5: Call the radar flow measurement model based on the Doppler effect to correct the preliminary water level prediction value sequence output in step S4. Through the preset water flow-water level coupling relationship in the model, compare and analyze the preliminary prediction value with the water flow velocity parameters in the original water flow characteristic parameter set in step S1 to generate a water level correction coefficient, and correct the preliminary water level prediction value sequence based on the coefficient. Specifically, step S5 is a crucial step in correcting the preliminary water level prediction sequence. Its core involves calling a radar flow measurement model based on the Doppler effect. Through a pre-defined flow-level coupling relationship and combining it with flow velocity parameters from the original flow characteristic parameter set, water level correction coefficients are generated to correct the preliminary prediction sequence. The technical parameters of this step include the model parameters of the flow-level coupling relationship. These parameters are fitted based on a large amount of historical flow and water level data, reflecting the actual changes in water level under different flow velocities. By introducing flow velocity parameters for correction, some physical characteristics that the Transformer network framework might overlook during the prediction process can be compensated for, improving the accuracy of water level prediction. The significance of this step lies in optimizing the preliminary prediction results through the combination of the physical model and the data model, reducing prediction errors, and making the predicted water level values ​​closer to the actual situation.

[0027] In the specific implementation process, firstly, the dynamic change sequence of water flow velocity parameters is extracted from the original water flow characteristic parameter set obtained in step S1. The time interval of this sequence is consistent with the preliminary water level prediction value sequence output in step S4, both being 5 minutes or 10 minutes, ensuring that the two sequences correspond one-to-one in the time dimension. Then, the water flow-water level coupling relationship module in the radar flow measurement model based on the Doppler effect is called. This module has a preset correlation equation between water flow velocity and water level obtained by fitting historical data. The extracted water flow velocity parameter sequence is input into this correlation equation to calculate the corresponding theoretical water level value sequence. For example, when the water flow velocity is 2 m / s, the theoretical water level value is 145 m; when the water flow velocity is 3 m / s, the theoretical water level value is 146.5 m, etc. Next, the deviation value between the preliminary water level prediction value sequence and the theoretical water level value sequence is calculated. For example, if the preliminary prediction value is 145.5 m and the theoretical water level value is 145 m, the deviation value is 0.5 m. The deviation values ​​at all time points are arranged chronologically to form a deviation sequence. A sliding window algorithm with a window size of 10 is used to smooth the deviation sequence, eliminating the influence of random errors, resulting in smoothed deviation values. A water level correction coefficient is calculated based on the smoothed deviation values. The formula for the correction coefficient is (1 + smoothed deviation value / theoretical water level value). For example, if the smoothed deviation value is 0.3 meters and the theoretical water level value is 145 meters, then the correction coefficient is 1 + 0.3 / 145 ≈ 1.002. Finally, each predicted value in the initial water level prediction sequence is multiplied by the corresponding correction coefficient to obtain the corrected water level prediction sequence. The accuracy of the corrected prediction can be improved to within ±0.3 meters.

[0028] Step S6: Perform time-series integration on the corrected water level prediction value sequence in Step S5 to obtain the water level change curve of the hydropower station in the monitoring area for a continuous time period. Each data point in the curve includes the corresponding timestamp and water level value, and the density of the data points matches the spatiotemporal sampling interval in Step S2.

[0029] Specifically, step S6 is the final stage of the water level monitoring process. Its main task is to integrate the corrected water level prediction value sequence over time to obtain the water level change curve of the hydropower station within the monitoring area for a continuous time period. The technical parameters of this step include the density of data points, which matches the spatiotemporal sampling interval in step S2 to ensure that the integrated water level change curve can completely and accurately reflect the changes in water level in time and space. Too high a data point density will result in an overly complex curve, hindering observation and analysis; too low a density may lead to the loss of important water level change information. The significance of this step lies in transforming scattered water level prediction values ​​into a continuous, visualized water level change curve, allowing staff to intuitively understand the dynamic trend of water level changes at the hydropower station, providing clear and accurate data for the operation scheduling and safety monitoring of the hydropower station.

[0030] In the specific implementation process, the time range for time series integration is first determined. This range is consistent with the time range of the original data collection in step S1. For example, if the original data collected 24 hours of water flow information, then the time range for time series integration is also 24 hours. Then, the data point density of the integrated water level change curve is determined according to the spatiotemporal sampling interval in step S2. If the spatiotemporal sampling interval is 5 minutes × 50 square meters, then the time interval between data points is 5 minutes, meaning there is one corresponding water level data point every 5 minutes. The corrected water level prediction value sequence is arranged according to the order of the timestamps to ensure the accuracy of the time order of each data point. For water level data points corresponding to subsets with spatially overlapping areas, a weighted average method is used for fusion. The weights are determined based on the area proportion of the subsets in the overlapping area. For example, if the area proportions of two subsets in an overlapping area are 60% and 40% respectively, then the weights of the corresponding water level data points are 0.6 and 0.4 respectively. The fused water level value is the weighted sum of the two data points. During the integration process, detailed information is added to each data point, including a timestamp (accurate to the second) and a water level value (accurate to 0.01 meters). Finally, data visualization tools were used to connect the integrated water level data points into a continuous water level change curve. The horizontal axis of the curve represents time, and the vertical axis represents the water level value. The range of the vertical axis was set from the dead water level of 120 meters to the design flood level of 160 meters, with a scale interval of 1 meter, to clearly display the water level changes. The integrated water level change curve was stored in a database and could be transmitted in real time to the monitoring terminal of the hydropower station for staff to view and analyze.

[0031] Preferably, in step S3, the encoding layer of the water level detection model based on the Transformer network framework also introduces historical monitoring parameters of the hydropower station's water level during the process of generating the water flow feature encoding vector. These historical monitoring parameters include the peak water level, valley water level, and water level change rate within a preset time period. The output of the multi-head self-attention mechanism is optimized using the following model formula: ; in, This is the optimized water flow feature encoding vector. This is the initial encoding vector output by the multi-head self-attention mechanism. This is the historical parameter influence factor, with a value range of 0.1-0.3. This represents the historical water level value for the same period at the current moment. This is the average water level value over a preset time period in the past. This refers to the highest water level value within a previously preset time period. This is the lowest water level value within a preset time period in the past.

[0032] Specifically, the core of the process for generating flow feature encoding vectors in the encoding layer of the water level detection model based on the Transformer network framework lies in introducing historical monitoring parameters of the hydropower station's water level. These historical monitoring parameters include the peak, trough, and rate of change of water level over a preset time period. These parameters provide important references for optimizing the encoding vectors. Regarding technical parameters, the historical parameter influence factor is set within the range of 0.1-0.3. This range is determined after comprehensively considering the influence of historical data on current predictions, avoiding excessive influence from historical data that could render the model insensitive to current flow changes, while ensuring effective correction of the encoding vectors. In implementation, water level data for a preset time period (e.g., the last 30 days) is first extracted from the hydropower station database. The average, highest, and lowest water level values ​​within this period are calculated, and the historical water level values ​​for the same period at the current moment are also obtained. These parameters are then substituted into the optimization process, and the optimized flow feature encoding vector is generated by adjusting the initial encoding vector output by the multi-head self-attention mechanism. The significance of this process lies in leveraging the periodic and trend patterns inherent in historical water level data to enhance the ability of the encoding vector to express the correlation between water flow characteristics and water level. This makes the encoding vector more closely match the actual water level change characteristics of the hydropower station, providing more accurate feature inputs for subsequent water level prediction, thereby improving the adaptability and prediction accuracy of the entire water level detection method under different hydrological cycles.

[0033] Preferably, in step S4, when the decoding layer of the Transformer network framework outputs the preliminary water level prediction sequence, it constrains the analysis process by incorporating the structural parameters of the hydropower station dam. These structural parameters include the dam height, the slope of the dam's upstream face, and the density of the dam's drainage outlets. The preliminary water level prediction is calculated using the following model formula: ; in, These are preliminary water level predictions. For the parsing function of the decoding layer, This is a water flow feature encoding vector. This is the influence coefficient of structural parameters, with a value range of 0.05-0.2. This represents the slope value of the dam's upstream side. This refers to the distribution density of drainage outlets within the dam body, i.e., the number of drainage outlets per unit area. This refers to the height of the dam.

[0034] Specifically, when the Transformer network framework's decoding layer outputs the preliminary water level prediction sequence, structural parameters of the hydropower station dam are introduced to constrain the analytical process, aiming to improve the rationality and accuracy of the preliminary predictions. These structural parameters include dam height, upstream slope, and drainage outlet density. These parameters directly affect the interaction between the water flow and the dam, thus influencing water level changes. Among the technical parameters, the influence coefficient of the structural parameters ranges from 0.05 to 0.2. This range is determined based on measured data of the impact of different dam structures on water level, ensuring that the adjustment range of the structural parameters on the preliminary predictions is within a reasonable range. During implementation, specific structural parameters of the hydropower station dam are first collected, such as measured dam height, calculated upstream slope, and the number of drainage outlets per unit area. These parameters, along with the water flow feature encoding vector generated in step S3, are input into the decoding layer's analytical process. The decoding layer's analytical function processes the encoding vector and, combined with the structural parameters, calculates the preliminary water level prediction. The significance of this mechanism lies in taking into account the impact of the dam structure, a fixed factor, on the water level. For example, when the slope of the dam's upstream side is steep, the impact of the water flow on the water level is more significant. Densely distributed drainage outlets facilitate water discharge, thereby lowering the water level. By incorporating these structural characteristics into the prediction process, the preliminary water level prediction value is made more consistent with the actual hydrological environment of the hydropower station, reducing prediction deviations caused by ignoring the influence of the dam.

[0035] Preferably, in step S5, during the generation of the water level correction coefficient by the radar flow measurement model based on the Doppler effect, water temperature parameters and water sediment content parameters are introduced, and the water level correction coefficient is calculated using the following model formula: ; in, This is the water level correction factor. This is the temperature influence coefficient, with a value ranging from 0.01 to 0.03. The current water temperature, For standard reference water temperature, The value is the influence coefficient of sediment content, ranging from 0.02 to 0.05. The sediment content of water is the mass of sediment per unit volume of water.

[0036] Specifically, by introducing water temperature and sediment concentration parameters, the process of generating water level correction coefficients for a radar flow measurement model based on the Doppler effect was optimized. Water temperature and sediment concentration are important environmental factors affecting water flow characteristics and water level detection. Temperature changes lead to alterations in water density and flow resistance, while sediment concentration affects water viscosity and radar signal reflection characteristics. Technically, the temperature influence coefficient ranges from 0.01 to 0.03, and the sediment concentration influence coefficient ranges from 0.02 to 0.05. These ranges were determined through extensive experiments, reflecting the degree of influence of temperature changes and sediment concentration differences on water level correction, respectively. In implementation, sensors deployed in the monitoring area first collect real-time data on current water temperature and sediment concentration, while simultaneously determining a standard reference water temperature (typically 20℃). These parameters are then substituted into the calculation process for the correction coefficients, and combined with the comparison results between the water flow velocity parameters and the preliminary water level prediction values, the water level correction coefficients are generated. This coefficient is used to correct the preliminary water level prediction sequence. Its significance lies in making up for the shortcomings of traditional correction processes that ignore changes in the physical properties of water bodies. This makes the corrected water level values ​​more reflective of the actual water body conditions. For example, when the water temperature rises during the high-temperature season, leading to a decrease in density, or when the sediment content of the water body increases during the flood season, the correction coefficient can make targeted adjustments to the preliminary prediction values, thereby improving the accuracy and stability of water level detection.

[0037] Preferably, in step S2, when extracting the original water flow characteristic parameter set in segments, dynamic adjustments are also made in conjunction with the precipitation parameters of the watershed where the hydropower station is located. The precipitation parameters include precipitation intensity and precipitation duration, and the spatiotemporal sampling interval is determined by the following model formula: ; in, The adjusted spatiotemporal sampling interval, The initial spatiotemporal sampling interval, The value is a precipitation impact factor, ranging from 0.005 to 0.02. Precipitation intensity, which is the amount of precipitation per unit time. This refers to the duration of precipitation.

[0038] Specifically, the segmented extraction process of the original water flow characteristic parameter set in step S2 is dynamically adjusted in conjunction with the precipitation parameters of the watershed where the hydropower station is located, so that the spatiotemporal sampling interval can better adapt to the impact of precipitation on water flow. Precipitation parameters include precipitation intensity (precipitation amount per unit time) and precipitation duration. These two parameters directly determine the severity of water flow changes. When the precipitation intensity is high and the duration is long, the changes in water flow velocity and direction are more frequent. Among the technical parameters, the precipitation influence factor ranges from 0.005 to 0.02. This range is determined based on the measured data of precipitation disturbance to water flow, ensuring that the adjustment of the sampling interval can capture water flow changes caused by precipitation in a timely manner without causing excessive data volume due to too short an interval. In practice, real-time precipitation data, including precipitation amount per minute and the duration of precipitation, is obtained through meteorological monitoring stations in the watershed. Based on this data, the adjusted spatiotemporal sampling interval was calculated. For periods of high precipitation intensity and long duration, the sampling interval was shortened to collect water flow characteristic parameters more intensively; for periods with no precipitation or weak precipitation, the sampling interval was maintained or extended. The significance of this dynamic adjustment mechanism is that it enables the segmented extracted water flow characteristic subsets to more accurately reflect the dynamic changes of water flow under different precipitation conditions, avoiding the problem of losing key information during precipitation or data redundancy during periods without precipitation caused by fixed sampling intervals. This provides more targeted input data for subsequent model processing.

[0039] Preferably, in step S6, when integrating the corrected water level prediction sequence, the operating parameters of the hydropower station generator units are introduced. These operating parameters include the generator power output and the number of operating units. The integrated water level change curve is then further optimized using the following model formula: ; in, This is the water level value after secondary optimization. This is the water level value after time-series integration. This is the unit impact factor, with a value ranging from 0.01 to 0.03. This refers to the power generation capacity of a single generating unit. Number of operating units This represents the maximum total power generation capacity of the hydropower station units.

[0040] Specifically, in step S6, when integrating the corrected water level prediction sequence, the operating parameters of the hydropower station's generator units are introduced to perform secondary optimization on the integrated water level change curve, reflecting the impact of unit operation on the water level. The generator unit operating parameters include the unit's power generation and the number of operating units. These parameters determine the water intake of the turbines, thus affecting water level changes. Among the technical parameters, the unit influence coefficient ranges from 0.01 to 0.03. This range is determined based on measured data of water level changes under different unit operating conditions, reasonably quantifying the degree of influence of unit operation on the water level. During implementation, real-time operating parameters are obtained from the hydropower station's unit monitoring system, including the current power generation of each operating unit and the number of operating units, while also determining the maximum total power generation of the hydropower station's units. These parameters are then substituted into the secondary optimization process to adjust the integrated water level values, generating the secondary optimized water level values. The significance of this optimization lies in taking into account the impact of the hydropower station's own operational activities on the water level. For example, when the generating units are operating at full load and the water intake is large, the water level will drop faster. Through secondary optimization, the water level change curve can more accurately reflect the water level fluctuations caused by the unit's operation, thereby providing water level data that is more in line with the actual operating conditions for the operation and scheduling of the hydropower station, and improving the scientificity and rationality of scheduling decisions.

[0041] Preferably, step S3 includes the following sub-steps: S31, performing parameter dimension alignment on the water flow feature subset obtained in step S2, transforming different types of water flow feature parameters to the same data dimension space, so that the parameters of water flow velocity, water flow direction, and water flow disturbance frequency have a unified dimension, wherein the dimension transformation process determines the transformation coefficient based on the distribution range of each parameter in historical data; S32, inputting the dimension-aligned water flow feature subset into the embedding layer of the encoding layer of the Transformer network framework in chronological order, and converting each parameter value into a high-dimensional vector representation through the embedding layer, wherein the vector dimension corresponding to each parameter is... The degree is determined based on the dynamic range of parameter changes; the larger the range of changes, the higher the corresponding vector dimension. In S33, the high-dimensional vector output from the embedding layer is input into the multi-head self-attention mechanism. This mechanism calculates the attention weights between different parameter vectors through multiple parallel attention heads. Each attention head focuses on different types of correlations between parameters, including linear and non-linear correlations. In S34, the multiple attention weight matrices output by the multi-head self-attention mechanism are weighted and summed to obtain a comprehensive attention weight matrix. This matrix is ​​then multiplied with the high-dimensional vector output from the embedding layer to generate a water flow feature encoding vector that includes parameter correlation weights.

[0042] Specifically, step S3 is refined, specifying four sub-steps to improve the quality of the water flow feature encoding vector through a more detailed processing flow. Step S31 performs parameter dimension alignment, transforming different types of parameters such as water flow velocity, direction, and disturbance frequency to the same data dimension space to ensure uniformity of units. The technical parameters for this process include conversion coefficients, which are determined based on the distribution range of each parameter in historical data. For example, historical water flow velocity data is distributed between 0.1-10 m / s, and water flow direction between 0-360 degrees. A mapping formula is used to convert both to a numerical range of 0-1, and the conversion coefficients are calculated based on the maximum and minimum values ​​of the distribution range. Step S32 inputs the dimension-aligned subset of data into the embedding layer, converting it into a high-dimensional vector representation. The vector dimension is related to the dynamic range of parameter changes; parameters with larger ranges of change correspond to higher dimensions. For example, water flow velocity has a large range of change and can be set to 128 dimensions, while water flow direction has a relatively small range of change and can be set to 64 dimensions. In step S33, the multi-head self-attention mechanism calculates attention weights between parameter vectors using multiple parallel attention heads. The number of attention heads is adjusted according to data complexity, typically ranging from 8 to 32. Each attention head focuses on different types of relationships, such as linear or non-linear relationships. The attention weights are obtained by multiplying the query, key, and value matrices. Step S34 performs a weighted summation of the multiple attention weight matrices, with weights allocated according to the importance of each attention head. This summation is then multiplied by the high-dimensional vector output from the embedding layer to generate a water flow feature encoding vector, typically with dimensions of 512 or 1024. The significance of this series of sub-steps lies in transforming the original water flow parameters into encoding vectors that better reflect the inherent patterns through dimensional unification, high-dimensional mapping, multi-dimensional association mining, and weight integration. This provides high-quality input for subsequent decoding processes and enhances the model's ability to capture complex water flow features. During implementation, it is crucial to strictly adhere to the parameter settings and processing logic of each step to ensure that the output of each step meets the input requirements of subsequent steps. Through layer-by-layer processing, the correlation information between water flow features and water level is gradually extracted.

[0043] Preferably, step S4 includes the following sub-steps: S41, concatenating the water flow feature encoding vector generated in step S3 with the preset hydropower station water level reference parameter vector to form the input vector of the decoding layer, wherein the water level reference parameter vector includes vector representations of the hydropower station's normal storage water level, dead water level, and design flood level parameters; S42, inputting the concatenated input vector into the first feedforward neural network layer of the decoding layer of the Transformer network framework, which performs feature mapping on the input vector through linear transformation and nonlinear activation functions to enhance the feature signals related to water level prediction in the vector; S43, inputting the output of the first feedforward neural network layer into the residual connection and layer normalization module of the decoding layer, retaining the original features of the input vector through residual connection, and adjusting the distribution of the feature vector through layer normalization to make the mean and variance of each element in the vector within a preset range; S44, inputting the feature vector after residual connection and layer normalization processing into the second feedforward neural network layer of the decoding layer, which further extracts features from the feature vector through multiple sets of convolutional kernels and outputs a preliminary water level prediction value sequence.

[0044] Specifically, step S4 was refined into four sub-steps to improve the accuracy of the preliminary water level prediction sequence. Step S41 concatenates the water flow feature encoding vector with the water level reference parameter vector. The water level reference parameters include normal storage level, dead water level, and design flood level, etc., and need to be converted into vectors of the same dimension as the encoding vector. For example, if the encoding vector is 512-dimensional, the reference parameter vector is also set to 512-dimensional, resulting in a 1024-dimensional input vector. Step S42 inputs the concatenated vector into the first feedforward neural network layer, which contains 2048 neurons. A linear transformation matrix (1024×2048 dimension) maps the input vector to 2048 dimensions, followed by ReLU activation to enhance effective feature signals and suppress noise. Step S43's residual connection adds the input vector to the feedforward layer output, preserving the original features. Layer normalization standardizes the vector, making the mean of each element 0 and the variance 1, stabilizing the feature distribution and avoiding gradient vanishing or exploding problems during training. Step S44 inputs the processed vector into the second feedforward neural network layer, which contains 1024 neurons. This layer outputs a 1-dimensional water level prediction value through a 1024×1 linear transformation matrix. Simultaneously, multiple sets of 1×1 convolutional kernels are used to further extract features. The number of convolutional kernels is related to the feature dimension, typically 64 or 128 sets. The significance of these four steps lies in transforming the feature information in the encoded vector into specific water level prediction values ​​by fusing baseline parameters, feature mapping, stable distribution, and final prediction, fully utilizing prior knowledge and data features to improve prediction accuracy. During implementation, it is crucial to ensure that the parameters of each layer are set reasonably. For example, the number of neurons in the neural network and the size of the convolutional kernels need to be verified and adjusted based on historical data. The calculations of residual connections and layer normalization must be accurate to guarantee the processing effect of each step. The final output sequence of preliminary water level prediction values ​​must correspond one-to-one with the time nodes of the subset dataset, with the error controlled within a preset range.

[0045] Preferably, step S5 includes the following sub-steps: S51, extracting the dynamic change sequence of water flow velocity parameters from the original water flow characteristic parameter set obtained in step S1, and synchronizing this sequence with the preliminary water level prediction value sequence output in step S4 to ensure that the data points in the two sequences correspond one-to-one in the time dimension; S52, calling the water flow-water level coupling relationship module in the radar flow measurement model based on the Doppler effect, which establishes a correlation model between water flow velocity and water level according to a preset physical equation, inputting the synchronized water flow velocity parameters into the correlation model to obtain the corresponding theoretical water level value sequence; S5 3. Calculate the deviation between the preliminary water level prediction value sequence output in step S4 and the theoretical water level value sequence obtained in step S52. Arrange the deviation values ​​in chronological order to form a deviation sequence, and smooth the deviation sequence using a sliding window algorithm to obtain the smoothed deviation value. S54. Calculate the water level correction coefficient based on the smoothed deviation value. Specifically, the smoothed deviation value at each time point is compared with the theoretical water level value at that time point to obtain the correction coefficient for that time point. The correction coefficients at all time points are combined into a water level correction coefficient sequence, which is used to correct the preliminary water level prediction value sequence.

[0046] Specifically, step S5 is refined into four sub-steps to improve the accuracy of water level correction. Step S51 extracts the dynamic sequence of water flow velocity from the original parameter set and synchronizes it with the preliminary water level prediction sequence to ensure complete consistency of timestamps. If there is a time deviation, it needs to be adjusted through interpolation or resampling. For example, if the time interval of the prediction sequence is 5 minutes, the water flow velocity sequence also needs to be adjusted to a 5-minute interval, using linear interpolation to fill in missing data. Step S52 calls the water flow-water level coupling module. This module establishes a correlation model based on physical equations, inputs the synchronized water flow velocity parameters, and outputs a theoretical water level value sequence. The model parameters are obtained by fitting historical water flow velocity and corresponding measured water level data, such as determining the functional relationship between velocity and water level through linear regression or nonlinear fitting. Step S53 calculates the deviation between the preliminary prediction value and the theoretical water level value, forming a deviation sequence, and then smooths it using a sliding window algorithm. The window size is set to 5-20 time points depending on the data fluctuation. The smoothed deviation value is obtained by calculating the average deviation within the window, eliminating the influence of random errors. Step S54 calculates the correction coefficient based on the ratio of the smoothed deviation value to the theoretical water level value. The correction coefficient for each time point is 1 plus (smoothed deviation value / theoretical water level value). The preliminary predicted value is then multiplied by the corresponding correction coefficient to obtain the corrected sequence. The significance of this series of steps lies in ensuring data comparability through time synchronization, using the theoretical water level obtained from the physical model as a reference, and achieving accurate correction of the preliminary predicted value through deviation analysis and correction coefficient calculation, thus integrating the advantages of data-driven models and physical models. During implementation, it is necessary to ensure the time synchronization accuracy between the flow velocity sequence and the predicted sequence, the parameters of the coupling model need to be updated periodically based on new measured data, the sliding window size needs to be determined experimentally to achieve the optimal value, and the calculation of the correction coefficient must be accurate to ensure that the corrected water level prediction value is closer to the actual water level, with the error reduced by more than 30% compared to the preliminary prediction.

[0047] The radar flow measurement model based on the Doppler effect is the key model in this invention for capturing water flow characteristics and correcting water level predictions. Specifically, it uses multiple radar flow measurement devices deployed in the monitoring area of ​​the hydropower station to continuously capture dynamic signals of water flow using the Doppler effect, thereby obtaining raw characteristic parameters such as flow velocity, direction, and disturbance frequency. This model pre-determines a flow-water level coupling relationship, establishing a correlation between parameters such as flow velocity and water level. Its main functions are twofold: first, in the data acquisition phase, it provides comprehensive raw flow characteristic parameters for the entire detection process, which form the basis for subsequent analysis and prediction; second, in the water level correction phase, by comparing the preliminary water level prediction with the raw flow velocity parameters, it generates correction coefficients to correct the preliminary prediction, compensating for the influence of physical characteristics that other models may overlook during the prediction process. This model is of great significance because it fully utilizes the advantages of the Doppler effect in dynamic signal detection, ensuring the accuracy and comprehensiveness of the raw water flow data. At the same time, by combining the physical model with the data model, it improves the accuracy and reliability of water level detection, making the detection results more consistent with the actual hydrological conditions of the hydropower station, and providing more solid data support for subsequent water level analysis and power station scheduling.

[0048] The water level detection model based on the Transformer network framework is the core model in this invention for deeply mining the correlation of water flow features and predicting water levels. Specifically, it is built on the Transformer network and includes an encoding layer and a decoding layer. The encoding layer uses a multi-head self-attention mechanism to mine the correlation between different parameters in the water flow feature subset, generating an encoded vector containing parameter correlation weights. The decoding layer combines the hydropower station's water level benchmark parameters to parse the encoded vector and output a preliminary sequence of predicted water levels. The main function of this model is to perform deep processing on the segmented extracted water flow feature subset, transforming the original water flow parameters into feature vectors that reflect the inherent laws, and then using these feature vectors to predict water levels. Its significance lies in overcoming the limitations of traditional models in terms of insufficient multi-parameter fusion capabilities. Through the multi-head self-attention mechanism, it can effectively capture the complex correlations between different water flow parameters, improving the understanding and expression of water flow features, thereby increasing the accuracy of water level prediction. Meanwhile, the model can adapt to water flow conditions of varying complexity. By adaptively adjusting parameters such as the number of attention heads, it enhances its adaptability to different hydrological environments, providing a more intelligent and efficient prediction method for hydropower station water level detection, which helps improve the scientific nature and safety of power station operation.

[0049] like Figure 2 As shown, a hydropower station water level detection system includes: A multi-source Doppler radar signal acquisition unit, which is connected to multiple Doppler effect radar flow measurement devices deployed in the predetermined monitoring area of ​​the hydropower station, is used to receive and store the original water flow characteristic parameter set captured by the radar flow measurement devices; The spatiotemporal feature segmentation extraction unit is connected to the multi-source Doppler radar signal acquisition unit. It is used to segment the original water flow feature parameter set according to the preset spatiotemporal sampling interval to generate multiple water flow feature subsets. The Transformer encoding processing unit, connected to the spatiotemporal feature segmentation extraction unit, is used to mine the correlation between water flow feature subsets through a multi-head self-attention mechanism and output water flow feature encoding vectors. The preliminary water level prediction unit is connected to the Transformer encoding processing unit. It is used to analyze the water flow feature encoding vector in combination with the hydropower station water level reference parameters and output a preliminary water level prediction value sequence. The radar model correction unit is connected to the preliminary water level prediction unit and the multi-source Doppler radar signal acquisition unit, respectively. It is used to call the radar flow measurement model based on the Doppler effect to correct the preliminary water level prediction value sequence and generate the corrected water level prediction value sequence. The time-series integration output unit is connected to the radar model correction unit. It is used to integrate the corrected water level prediction value sequence in time and output the water level change curve of the hydropower station. This unit is also connected to the monitoring terminal of the hydropower station to transmit the water level change curve to the monitoring terminal for display.

[0050] A method and system for detecting water levels in hydropower stations is disclosed. The primary advantage of this method and system is its ability to deeply integrate multi-source hydrological parameters, effectively overcoming the shortcomings of traditional techniques in parameter correlation mining. By utilizing multiple sets of Doppler effect radar flow measurement devices, the system comprehensively captures raw features such as water flow velocity, direction, and disturbance frequency, extracting them into time-series subsets through spatiotemporal segmentation. Employing the multi-head self-attention mechanism of the Transformer network framework's encoding layer, the system deeply mines the correlations between different parameters, generating encoding vectors containing correlation weights. This fully utilizes the inherent connections between parameters, overcoming the problem of insufficient detection accuracy caused by existing technologies relying on single or simple parameter combinations, and providing a solid data foundation for accurate prediction.

[0051] Secondly, it possesses strong environmental adaptability, capable of handling complex operating conditions and overcoming the shortcomings of traditional technologies in terms of poor adaptability. The decoding layer combines the hydropower station's water level benchmark parameters with the parsed encoding vector, while also incorporating parameters such as historical water levels and dam structure to optimize the prediction process. During radar flow measurement model calibration, parameters such as water temperature and sediment content are incorporated to generate calibration coefficients, and the sampling interval can be dynamically adjusted based on precipitation parameters, integrating the results with unit operating parameters for optimization. This multi-dimensional parameter fusion and dynamic adjustment mechanism enables the system to maintain stable detection performance even when facing changes in the hydrological environment and differences in dam structure, avoiding the problem of existing technologies being susceptible to external interference.

[0052] Furthermore, the continuity and reliability of the detection results are significantly improved, providing strong support for power station operation. By integrating the corrected water level prediction sequence over time, a continuous change curve containing timestamps and water level values ​​is obtained. The data point density matches the sampling interval, fully reflecting the dynamic changes in water level. This continuous and accurate detection result overcomes the shortcomings of traditional technology in terms of delayed or fragmented results, providing timely and reliable basis for hydropower station scheduling decisions and safety control, helping to improve the efficiency and safety of power station operation, and meeting the water level monitoring needs under different operating conditions.

[0053] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting water level in a hydropower station, characterized in that, Includes the following steps: Step S1: Using multiple sets of Doppler effect radar flow measurement devices, continuously capture the dynamic water flow signals formed during the water flow process in the monitoring area, and obtain the original water flow characteristic parameter set including water flow velocity, water flow direction and water flow disturbance frequency. Step S2: Based on the preset spatiotemporal sampling interval, the obtained original water flow feature parameter set is segmented and extracted to obtain multiple water flow feature subsets with time series attributes. Each subset corresponds to a continuous spatiotemporal sampling unit, and each subset includes the dynamic change trajectory of the water flow feature parameters within the sampling unit. Step S3: Input the obtained water flow feature subset into the encoding layer of the pre-built water level detection model based on the Transformer network framework. The multi-head self-attention mechanism in the encoding layer is used to mine the correlation between different parameters in the water flow feature subset and generate a water flow feature encoding vector including parameter correlation weights. Step S4: Input the generated water flow feature encoding vector into the decoding layer of the Transformer network framework. The decoding layer combines the preset hydropower station water level reference parameters to parse the encoding vector layer by layer and output a preliminary water level prediction value sequence. Step S5: Call the radar flow measurement model based on the Doppler effect to correct the output preliminary water level prediction value sequence. Through the preset water flow-water level coupling relationship in the model, compare and analyze the preliminary prediction value with the water flow velocity parameters in the original water flow characteristic parameter set in step S1, generate water level correction coefficient, and correct the preliminary water level prediction value sequence according to the coefficient. Step S6: Perform time-series integration on the corrected water level prediction value sequence to obtain the water level change curve of the hydropower station in the monitoring area for a continuous time period. Each data point in the curve includes the corresponding timestamp and water level value.

2. The method for detecting water level in a hydropower station according to claim 1, characterized in that, In step S2, when extracting the original water flow characteristic parameter set in segments, dynamic adjustments are also made in conjunction with the precipitation parameters of the watershed where the hydropower station is located. The precipitation parameters include precipitation intensity and precipitation duration. The spatiotemporal sampling interval is determined by the following model formula: ； in, The adjusted spatiotemporal sampling interval, The initial spatiotemporal sampling interval, The value is a precipitation impact factor, ranging from 0.005 to 0.

02. Precipitation intensity, which is the amount of precipitation per unit time. This refers to the duration of precipitation.

3. The method for detecting water level in a hydropower station according to claim 1, characterized in that, In step S3, when the coding layer generates the water flow feature coding vector, it also introduces historical monitoring parameters of the hydropower station's water level. These historical monitoring parameters include the peak water level, valley water level, and water level change rate within a preset time period. The output of the multi-head self-attention mechanism is optimized using the following model formula: ； in, This is the optimized water flow feature encoding vector. This is the initial encoding vector output by the multi-head self-attention mechanism. This is the historical parameter influence factor, with a value range of 0.1-0.

3. This represents the historical water level value for the same period at the current moment. This is the average water level value over a preset time period in the past. This refers to the highest water level value within a preset time period in the past. This is the lowest water level value within a preset time period in the past.

4. The method for detecting water level in a hydropower station according to claim 1, characterized in that, In step S4, when the decoding layer of the Transformer network framework outputs the preliminary water level prediction sequence, it constrains the analysis process by incorporating the structural parameters of the hydropower station dam. These structural parameters include the dam height, the slope of the dam's upstream face, and the density of the dam's drainage outlets. The preliminary water level prediction is calculated using the following model formula: ； in, These are preliminary water level predictions. For the parsing function of the decoding layer, This is a water flow feature encoding vector. This is the influence coefficient of structural parameters, with a value range of 0.05-0.

2. This represents the slope value of the dam's upstream side. This refers to the distribution density of drainage outlets within the dam body, i.e., the number of drainage outlets per unit area. This refers to the height of the dam.

5. The method for detecting water level in a hydropower station according to claim 1, characterized in that, In step S5, when the radar flow measurement model generates the water level correction coefficient, water temperature parameters and water sediment content parameters are introduced, and the water level correction coefficient is calculated using the following model formula: ； in, This is the water level correction factor. This is the temperature influence coefficient, with a value ranging from 0.01 to 0.

03. The current water temperature, For standard reference water temperature, The value is the influence coefficient of sediment content, ranging from 0.02 to 0.

05. The sediment content of water is the mass of sediment per unit volume of water.

6. The method for detecting water level in a hydropower station according to claim 1, characterized in that, In step S6, when integrating the corrected water level prediction sequence over time, the operating parameters of the hydropower station generator units are introduced. These operating parameters include the generator power output and the number of operating units. The integrated water level change curve is then further optimized using the following model formula: ； in, This is the water level value after secondary optimization. This is the water level value after time-series integration. This is the unit impact factor, with a value ranging from 0.01 to 0.

03. This refers to the power generation capacity of a single generating unit. Number of operating units This represents the maximum total generating capacity of the hydropower station's generating units.

7. A method for detecting water level in a hydropower station according to claim 1 or 3, characterized in that, Step S3 includes the following sub-steps: S31, performing parameter dimension alignment on the water flow feature subset obtained in step S2, transforming different types of water flow feature parameters to the same data dimension space, so that the parameters of water flow velocity, water flow direction, and water flow disturbance frequency have a unified dimension. The dimension transformation process determines the transformation coefficients based on the distribution range of each parameter in historical data; S32, inputting the dimension-aligned water flow feature subset into the embedding layer of the Transformer network framework's encoding layer in chronological order. The embedding layer converts each parameter value into a high-dimensional vector representation, where the root dimension of the vector corresponding to each parameter is... The dynamic range of parameter changes determines the vector dimension; the larger the range, the higher the corresponding vector dimension. In S33, the high-dimensional vector output from the embedding layer is input into the multi-head self-attention mechanism. This mechanism calculates the attention weights between different parameter vectors through multiple parallel attention heads. Each attention head focuses on different types of correlations between parameters, including linear and non-linear correlations. In S34, the multiple attention weight matrices output by the multi-head self-attention mechanism are weighted and summed to obtain a comprehensive attention weight matrix. This matrix is ​​then multiplied with the high-dimensional vector output from the embedding layer to generate a water flow feature encoding vector that includes parameter correlation weights.

8. A method for detecting water level in a hydropower station according to claim 1 or 4, characterized in that, Step S4 includes the following sub-steps: S41, concatenating the water flow feature encoding vector generated in step S3 with the preset hydropower station water level reference parameter vector to form the input vector of the decoding layer, wherein the water level reference parameter vector includes vector representations of the hydropower station's normal storage water level, dead water level, and design flood level parameters; S42, inputting the concatenated input vector into the first feedforward neural network layer of the Transformer network framework decoding layer, which performs feature mapping on the input vector through linear transformation and nonlinear activation functions to enhance the feature signals related to water level prediction in the vector; S43, inputting the output of the first feedforward neural network layer into the residual connection and layer normalization module of the decoding layer, retaining the original features of the input vector through residual connection, and adjusting the distribution of the feature vector through layer normalization to ensure that the mean and variance of each element in the vector are within a preset range; S44, inputting the feature vector after residual connection and layer normalization processing into the second feedforward neural network layer of the decoding layer, which further extracts features from the feature vector through multiple sets of convolutional kernels and outputs a preliminary water level prediction sequence.

9. A method for detecting water level in a hydropower station according to claim 1 or 5, characterized in that, Step S5 includes the following sub-steps: S51, extracting the dynamic change sequence of water flow velocity parameters from the original water flow characteristic parameter set obtained in step S1, and synchronizing this sequence with the preliminary water level prediction value sequence output in step S4 to ensure that the data points in the two sequences correspond one-to-one in the time dimension; S52, calling the water flow-water level coupling relationship module in the radar flow measurement model based on the Doppler effect, which establishes a correlation model between water flow velocity and water level according to preset physical equations, inputting the synchronized water flow velocity parameters into the correlation model to obtain the corresponding theoretical water level value sequence; S53, The deviation between the preliminary water level prediction value sequence output in step S4 and the theoretical water level value sequence obtained in step S52 is calculated. The deviation value is arranged in chronological order to form a deviation sequence. The deviation sequence is then smoothed using a sliding window algorithm to obtain a smoothed deviation value. In step S54, the water level correction coefficient is calculated based on the smoothed deviation value. Specifically, the smoothed deviation value at each time point is compared with the theoretical water level value at that time point to obtain the correction coefficient for that time point. The correction coefficients at all time points are combined into a water level correction coefficient sequence, which is used to correct the preliminary water level prediction value sequence.

10. A hydropower station water level detection system, characterized in that, include: A multi-source Doppler radar signal acquisition unit, which is connected to multiple Doppler effect radar flow measurement devices deployed in the predetermined monitoring area of ​​the hydropower station, is used to receive and store the original water flow characteristic parameter set captured by the radar flow measurement devices; The spatiotemporal feature segmentation extraction unit is connected to the multi-source Doppler radar signal acquisition unit. It is used to segment the original water flow feature parameter set according to the preset spatiotemporal sampling interval to generate multiple water flow feature subsets. The Transformer encoding processing unit, connected to the spatiotemporal feature segmentation extraction unit, is used to mine the correlation between water flow feature subsets through a multi-head self-attention mechanism and output water flow feature encoding vectors. The preliminary water level prediction unit is connected to the Transformer encoding processing unit. It is used to analyze the water flow feature encoding vector in combination with the hydropower station water level reference parameters and output a preliminary water level prediction value sequence. The radar model correction unit is connected to the preliminary water level prediction unit and the multi-source Doppler radar signal acquisition unit, respectively. It is used to call the radar flow measurement model based on the Doppler effect to correct the preliminary water level prediction value sequence and generate the corrected water level prediction value sequence. The time-series integration output unit is connected to the radar model correction unit. It is used to integrate the corrected water level prediction value sequence in time and output the water level change curve of the hydropower station. This unit is also connected to the monitoring terminal of the hydropower station to transmit the water level change curve to the monitoring terminal for display.

Images