Reservoir abnormal monitoring data judgment method and system

Abstract

The invention relates to a reservoir abnormal monitoring data judgment method and system, and belongs to the technical field of reservoir monitoring and data processing. According to the method, firstly, a filtering rule base based on physical parameter rules and instrument state basic threshold values is constructed, initial abnormal data which do not conform to the physical rules and are caused by instrument faults are quickly eliminated, and effective data are output; and then fine judgment based on W-RMSE and a brain-like pulse network SNN is executed on the effective data, the W-RMSE introduces a working condition and data credibility weight coefficient to adapt to extreme working condition fluctuation, the SNN captures time sequence data features through an LIF neuron model, and finally the results of the two are fused to output final judgment. By means of the method, the redundant data processing amount can be reduced by 30%-50%, the judgment precision reaches 92% or above, the data anomaly judgment efficiency and reliability under the extreme working condition are remarkably improved, and high-quality data support is provided for reservoir dispatching decision making.

Description

Technical Field

[0001] This invention relates to a method and system for determining abnormal monitoring data in reservoirs, belonging to the field of reservoir monitoring and data processing technology. Background Technology

[0002] Reservoir operation may encounter special hydrological and meteorological scenarios, including but not limited to torrential rain (such as red rainstorm warnings), typhoons, floods (such as inflow exceeding warning levels), and extreme low water levels (such as water levels approaching dead storage levels). Under these scenarios, the normal fluctuation range of reservoir monitoring data will significantly expand, and it is easily confused with abnormal data caused by instrument malfunctions, a phenomenon known as extreme conditions. Reservoir monitoring data (such as water level, rainfall, flow rate, and seepage pressure) is the core basis for the safe operation and scientific scheduling of reservoirs. Especially under extreme conditions, the authenticity and validity of the data directly determine the accuracy of scheduling decisions. In existing technologies, the determination of anomalies in reservoir monitoring data under extreme conditions generally adopts a process of "multi-source data cross-validation + model inference." Specifically, this involves directly performing multi-dimensional verification (such as spatiotemporal consistency verification and historical data comparison) on all collected monitoring data, and then using traditional machine learning models (such as LSTM and support vector machines) or error analysis methods (such as RMSE) to determine whether the data is abnormal.

[0003] However, existing technologies have two major drawbacks: 1. Failure to remove initial anomalous data leads to wasted computing power and interference with judgments: Under extreme conditions, monitoring instruments are prone to momentary malfunctions (such as sensor short circuits or signal transmission interruptions), generating "obviously unreasonable" initial anomalous data. Examples include negative rainfall monitoring values, water level monitoring values ​​far exceeding the reservoir's physical limits (exceeding the dam crest elevation or falling below the dead water level), and sudden rises and falls in seepage pressure monitoring values ​​without any hydrological logic support. Existing technologies directly process datasets containing such data. On the one hand, redundant initial anomalous data consumes significant computing resources, reducing data processing efficiency; on the other hand, initial anomalous data severely interferes with error calculation and model inference. For example, it inflates RMSE calculation results, causing normal data to be misjudged as anomalous; or it contaminates model training features, preventing the model from accurately learning the fluctuation patterns of normal data under extreme conditions, further reducing judgment accuracy.

[0004] 2. A single judgment method cannot take into account the complexity of extreme conditions: Under extreme conditions, the normal fluctuation range of reservoir monitoring data will be significantly expanded (such as the rapid rise of water level during flood season). Relying solely on RMSE cannot distinguish between "normal large fluctuations" and "abnormal fluctuations caused by instrument failure". Traditional machine learning models are poorly adapted to small sample data under extreme conditions and have difficulty accurately capturing the characteristics of data anomalies, resulting in insufficient reliability of the judgment results. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of low efficiency and poor accuracy in determining anomalies in reservoir monitoring data under extreme conditions in the prior art, and to provide a method for determining anomalies in reservoir monitoring data.

[0006] The technical solution adopted in this invention is as follows: A method for determining abnormal monitoring data in reservoirs includes the following steps: S1. Construct a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data, and filter out initial abnormal data that do not conform to the rules; S2. Construct a brain-like spiking network (SNN), using the filtered monitoring data time series, the corresponding hydrological parameters, and the instrument operating status parameters as inputs, setting hidden layers, and training the SNN anomaly recognition model with data anomaly classification as the output. S3. Construct a conventional anomaly prediction model to obtain a predicted value as a reference value, introduce "operating condition weight coefficient" and "data credibility weight coefficient", construct W-RMSE, and set the W-RMSE threshold to make a preliminary judgment on data anomalies; S4. The final judgment result is obtained by fusing the SNN anomaly identification result and the W-RMSE preliminary judgment result; S5. Based on the judgment results and manually reviewed data, regularly update the filtering rule base and SNN model parameters.

[0007] In the above method, the physical parameter constraints mentioned in step S1 include the setting of thresholds for water level, rainfall, flow rate, and seepage pressure.

[0008] The inputs in step S2 include water level, rainfall, rainfall level, inflow rate, instrument signal strength, and power supply voltage; the outputs include three types of judgment results: "normal data", "instrument malfunction abnormal data", and "extreme working condition unconventional data".

[0009] The formula for calculating W-RMSE in step S3 is: , in: This is the i-th actual monitoring data; The reference value for the i-th data is generated by an LSTM baseline model trained on historical extreme working condition data from the same period. The working condition weighting coefficient is set based on the intensity of the current extreme working condition. The stronger the working condition, the higher the weighting coefficient. The larger; Data credibility weighting coefficient: set based on the instrument's real-time operating status; n is the number of valid data items after filtering.

[0010] Set W-RMSE threshold (Set according to parameter type, such as water level data) ),like The initial assessment is that it is "normal data"; if The data was initially identified as "suspected abnormal data".

[0011] The rules for fusion judgment in step S4 are as follows: The confidence threshold for the SNN model classification results is T2. If the probability of the SNN model classifying a certain category is ≥ T2, the result of the SNN model is directly accepted. Otherwise, the classification results are fused according to weights. The classification accuracy and recall of W-RMSE and SNN are calculated through historical sample sets. The weights are assigned using F1 scores, and the two classification results are fused according to the weights.

[0012] Another object of the present invention is to provide a reservoir anomaly monitoring data determination system, comprising: The data acquisition module collects monitoring data and instrument operating status data in real time; The filtering rule base construction module is used to build a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data; The initial abnormal data filtering module filters out initial abnormal data that does not conform to the rules; The SNN anomaly detection model construction and anomaly detection module are used to construct a brain-like spiking network (SNN). The filtered monitoring data time series, corresponding hydrological parameters, and instrument operating status parameters are used as inputs. Hidden layers are set, and the SNN anomaly detection model is trained with data anomaly classification as the output, and anomaly detection is performed. Anomaly prediction module: responsible for generating reference values ​​required for W-RMSE calculation. ; The W-RMSE calculation module and anomaly identification module introduce "operating condition weight coefficient" and "data credibility weight coefficient" to calculate W-RMSE and set W-RMSE thresholds for preliminary judgment of data anomalies; The fusion judgment module fuses the SNN anomaly identification results and the W-RMSE preliminary judgment results to obtain the final judgment result; The optimization and update module regularly updates the filtering rule base and SNN model parameters based on the judgment results and manually reviewed data.

[0013] The system also includes a result output and alarm module, which outputs the final results to the scheduling platform and triggers alarms for abnormal instrument fault data.

[0014] The aforementioned anomaly prediction module comprises three sub-modules. Data preprocessing submodule: Cleans (removes initial abnormal data) and normalizes (standardizes to the [0,1] range according to parameter type) the historical extreme working condition data (last 5 years); The baseline model training submodule uses a lightweight LSTM network. The input is "the monitoring time series of the past 6 hours + rainfall level + inflow rate", and the output is the predicted value after 1 hour. The training process uses Early Stopping (patience=10) to avoid overfitting. Real-time prediction submodule: Receives filtered real-time monitoring data, calls the trained baseline model, and outputs reference values ​​for the corresponding time point. The prediction delay is ≤1 minute.

[0015] The beneficial effects of this invention are: 1. Improve judgment efficiency: Through the initial abnormal data filtering process, 30%-50% of redundant abnormal data can be quickly removed (based on actual reservoir monitoring data statistics), which greatly reduces the amount of data processed by the subsequent weighted RMSE calculation and SNN model, reduces computing power consumption, and improves processing efficiency by more than 40%. 2. Improved judgment accuracy: The removal of initial abnormal data avoids its interference with RMSE calculation and SNN model inference. Combining the advantages of weighted fusion RMSE (adapting to extreme operating condition fluctuations) and SNN (accurately distinguishing data types), the accuracy of distinguishing between "instrument failure abnormal data" and "extreme operating condition unconventional data" can reach over 92%, which is 15%-20% higher than existing technologies. 3. Enhanced practicality: The system can output judgment results in real time and trigger instrument fault alarms. It also supports dynamic optimization of rules and models, adapting to different reservoir monitoring scenarios and extreme working conditions, making it highly practical. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the SNN model structure of the present invention. Detailed Implementation

[0017] The following description, in conjunction with specific embodiments, provides further details.

[0018] Example 1: A method for determining abnormal monitoring data in a reservoir, comprising the following steps: S1. Construct a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data, and filter out initial abnormal data that does not conform to the rules: For different types of reservoir monitoring data (water level, rainfall, flow rate, seepage pressure), a filtering rule base based on physical parameter rules and instrument status thresholds is constructed to quickly filter out "obviously unreasonable" initial abnormal data, specifically including: Physical parameter rule filtering: Constraints are set based on the inherent physical parameters of the reservoir, for example, Water level data: Filter out monitoring values ​​that exceed the "dead water level - dam crest elevation" range; Rainfall data: Filter out negative values ​​and monitoring values ​​that exceed the instrument's maximum measurement range (e.g., 200 mm / h); Flow data: Negative values ​​and flow values ​​exceeding the maximum discharge capacity of the reservoir spillway are filtered out.

[0019] Instrument status filtering: Combine the real-time operating status data of the monitoring instrument (such as power supply voltage and signal strength) to filter out the monitoring data collected when the instrument is faulty (such as voltage below the threshold or signal interruption); if the instrument status data is missing, the data mutation characteristics are used to assist in the judgment. For example, if the difference between the monitoring value at a certain moment and the average value of the previous 10 minutes exceeds 3 times the standard deviation and there is no corresponding hydrological event (such as rainstorm warning) to support it, it is judged as initial abnormal data.

[0020] After filtering, only monitoring data that conforms to physical rules and is in normal instrument condition (or whose data fluctuations conform to preliminary hydrological logic) are retained for subsequent fine-tuning.

[0021] S2. Construct a brain-like spiking network (SNN), using the filtered monitoring data time series, corresponding hydrological parameters, and instrument operating status parameters as inputs, setting hidden layers, and training the SNN anomaly recognition model with data anomaly classification as the output: Brain-inspired spiking networks (SNNs) possess the ability to simulate the spiking firing of neurons in the human brain, exhibiting strong capabilities in capturing dynamic features of time-series data. They are particularly suitable for monitoring complex fluctuations in data under extreme conditions. The specific steps for constructing the SNN anomaly detection model in this invention are as follows: Input layer: Input features include filtered monitoring data time series (such as water level / rainfall data in the last hour), corresponding hydrological parameters (such as rainfall level and inflow), and instrument operating status parameters (such as signal strength and power supply voltage). Hidden layer: The LIF (Leaky Integrate-and-Fire) neuron model is adopted. Data features are encoded by the pulse firing frequency. 2-3 hidden layers are set, and the number of neurons in each layer is adjusted according to the data type of the monitored data (e.g., 64 neurons for water level data and 32 neurons for rainfall data). Output layer: The output consists of three categories of judgment results: "normal data", "instrument failure abnormal data" and "extreme working condition unconventional data". The Softmax activation function is used to output the probability value of each category.

[0022] The SNN model is trained using historical extreme operating condition data (including known instrument failure data and normal extreme operating condition data). During training, a cross-entropy loss function is used to optimize model parameters, ensuring the model's accuracy in distinguishing between "unusual extreme operating conditions" and "abnormal instrument failures." A probability threshold is set. (e.g., 0.8), if the model outputs the probability of a certain class. If the probability of all categories is equal, then it is initially determined to be data of that category; if the probability of all categories is equal... If so, it is initially determined to be "suspected abnormal data".

[0023] S3. Construct a conventional anomaly prediction model to obtain a predicted value as a reference value. Introduce "operating condition weight coefficient" and "data credibility weight coefficient" to construct W-RMSE. Set a W-RMSE threshold to make a preliminary judgment on data anomalies. Traditional RMSE calculations assign equal weight to all data points, failing to distinguish between "normal large fluctuations" and "abnormal fluctuations" under extreme operating conditions. This invention introduces "operating condition weight coefficients" and "data reliability weight coefficients" to construct a weighted fusion RMSE, with the specific formula as follows: , in: This is the i-th actual monitoring data; The reference value for the i-th data point is generated by an LSTM baseline model trained with historical extreme working condition data from the same period (the model input consists of historical monitoring time series and hydrological working condition parameters, with cross-validation RMSE ≤ 0.5). The operating condition weighting coefficient is set based on the intensity of the current extreme operating condition (such as rainstorm level, flood flow level). The stronger the operating condition, the higher the weighting coefficient. The larger the (e.g., a red alert for heavy rain) the better. Orange alert To reduce the impact of normal large fluctuations on RMSE under extreme operating conditions; Data credibility weighting coefficients: set based on the instrument's real-time operating status (good status = 1.0, average status = 0.6, poor status = 0.3). n is the number of valid data items after filtering.

[0024] Set W-RMSE threshold (Set according to parameter type, such as water level data) ),like The initial assessment is that it is "normal data"; if The data was initially identified as "suspected abnormal data".

[0025] S4. The final judgment result is obtained by fusing the SNN anomaly detection result and the W-RMSE preliminary judgment result: Construct fusion determination rules: The confidence threshold for the classification results of the SNN model is T2. If the probability of the SNN model in classifying a certain category is greater than or equal to T2, the result of the SNN model is directly accepted. In other cases, the results are fused according to the weights.

[0026] T2 is determined through cross-validation using historical labeled data (including samples of known "instrument malfunctions", "extreme operating conditions", and "normal data"). A historical labeled sample set (containing ≥1000 samples of each category) is selected, and candidate thresholds in the range of 0.7 to 0.9 are traversed. The optimal threshold is determined by using the maximum F1 score for "instrument malfunction data" and "extreme operating conditions" as the target. (Default value, can be adjusted according to the sample distribution of a specific reservoir).

[0027] The classification accuracy and recall of W-RMSE and SNN were calculated using historical sample sets, and the weights were assigned using F1 scores. The basis for the example weights (0.3 for W-RMSE, 0.7 for SNN) is as follows: Based on historical data verification, the SNN classification F1 score (0.91) for "instrument failure" and "extreme working conditions" is significantly higher than that for W-RMSE (0.75), so SNN is given a higher weight.

[0028] S5. Based on the judgment results and manually reviewed data, regularly update the filtering rule base and SNN model parameters: The final judgment results (normal data, abnormal data due to instrument failure, and unconventional data due to extreme conditions) are output to the reservoir dispatching platform in real time. An alarm signal is triggered for "abnormal data due to instrument failure" to prompt maintenance personnel to repair the corresponding instrument. At the same time, the judgment results are associated with and stored with the original data for subsequent optimization of the filtering rule base and SNN model parameters.

[0029] Taking actual monitoring data from a large reservoir (total capacity 1 billion m³, dam crest elevation 150 m, dead water level 100 m) as an example: (1) Data acquisition: The water level data (sampling frequency 1 time / 5 minutes), rainfall data (sampling frequency 1 time / 10 minutes), seepage pressure data (sampling frequency 1 time / 30 minutes), and the corresponding instrument operation status data (power supply voltage 12V±0.5V, signal strength ≥-80dBm) of the reservoir are collected in real time through the data acquisition module.

[0030] (2) Construct a filtering rule base and perform initial abnormal data filtering. Build a filtering rule base: Water level data: Filter out monitoring values ​​<100m or >150m; Rainfall data: Filter out monitoring values ​​of <0mm or >200mm / 10 minutes; Infiltration pressure data: Filter out monitoring values ​​that differ from the previous hour's mean by more than 3 standard deviations and are not supported by rainfall events; Instrument status: Data acquired when the power supply voltage is < 11.5V or the signal strength is < -80dBm is filtered out.

[0031] Suppose that among the water level data collected during a certain period, there are 120m (normal), 95m (below the dead water level, initially abnormal), 155m (above the dam crest, initially abnormal), and 130m (normal). After filtering, only the water level data of 120m and 130m are retained for the fine judgment stage.

[0032] (3) Calculation of W-RMSE: Reference value The model was generated from a baseline model trained using water level data from the reservoir during the past 10 years of flood seasons, assuming that 120m corresponds to... corresponding ; Operating condition weighting coefficient The current alert level is red for heavy rain (flood condition). ; Data credibility weighting coefficient The instrument is in normal condition (power supply 12.2V, signal -75dBm), settings are correct.

[0033] Calculate W-RMSE: , set up Since 1.56 ≤ 2.0, the water level data for this period is preliminarily determined to be "normal data".

[0034] (4) SNN model construction and anomaly prediction Input characteristics: Water level data for the past hour (12 data points: 115m, 117m, 118m, 120m, 122m, 125m, 127m, 128m, 129m, 130m, 131m, 132m), current rainfall (80mm / h, red alert), instrument status (power supply 12.2V, signal -75dBm); SNN model: There are 2 hidden layers, with 64 LIF neurons in the first layer and 32 LIF neurons in the second layer, and 3 neurons in the output layer (corresponding to the three types of results). Model output: "Normal data" probability 0.92, "Abnormal data due to instrument malfunction" probability 0.05, "Extreme operating condition unconventional data" probability 0.03; set up =0.8. Since the probability of "normal data" is 0.92≥0.8, it is initially determined to be "normal data".

[0035] (5) Comprehensive judgment The W-RMSE weight is set to 0.3 and the SNN weight to 0.7. Both are judged as "normal data". The final output is "normal data" and the result is uploaded to the reservoir scheduling platform.

[0036] If at a certain moment, the SNN model outputs a probability of 0.85 for "instrument malfunction and abnormal data," while the W-RMSE calculation result is... According to the fusion rules, the SNN result is given priority and is ultimately determined as "abnormal data due to instrument malfunction". An alarm signal is then triggered to prompt maintenance personnel to repair the corresponding water level gauge.

[0037] Example 2: A reservoir anomaly monitoring data determination system, comprising: The data acquisition module collects monitoring data and instrument operating status data in real time; The filtering rule base construction module is used to build a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data; The initial abnormal data filtering module filters out initial abnormal data that does not conform to the rules; The SNN anomaly detection model construction and anomaly detection module are used to construct a brain-like spiking network (SNN). The filtered monitoring data time series, corresponding hydrological parameters, and instrument operating status parameters are used as inputs. Hidden layers are set, and the SNN anomaly detection model is trained with data anomaly classification as the output, and anomaly detection is performed. Anomaly prediction module: responsible for generating reference values ​​required for W-RMSE calculation. ; The W-RMSE calculation module and anomaly identification module introduce "operating condition weight coefficient" and "data credibility weight coefficient" to calculate W-RMSE and set W-RMSE thresholds for preliminary judgment of data anomalies; The fusion judgment module fuses the SNN anomaly identification results and the W-RMSE preliminary judgment results to obtain the final judgment result; The optimization and update module regularly updates the filtering rule base and SNN model parameters based on the judgment results and manually reviewed data.

[0038] The system also includes a result output and alarm module, which outputs the final results to the scheduling platform and triggers alarms for abnormal instrument fault data.

[0039] The aforementioned anomaly prediction module comprises three sub-modules. Data preprocessing submodule: Cleans (removes initial abnormal data) and normalizes (standardizes to the [0,1] range according to parameter type) the historical extreme working condition data (last 5 years); The baseline model training submodule uses a lightweight LSTM network. The input is "the monitoring time series of the past 6 hours + rainfall level + inflow rate", and the output is the predicted value after 1 hour. The training process uses Early Stopping (patience=10) to avoid overfitting. Real-time prediction submodule: Receives filtered real-time monitoring data, calls the trained baseline model, and outputs reference values ​​for the corresponding time point. The prediction delay is ≤1 minute.

[0040] The sampling frequency of the data acquisition module is configurable. The sampling frequency for water level data is configured to be once every 5 minutes, for rainfall data to be once every 10 minutes, for seepage pressure data to be once every 30 minutes, and for flow rate data to be once every 15 minutes. The data acquisition module supports 4G, 5G, and LoRa communication protocols to achieve wireless data transmission with reservoir monitoring instruments.

[0041] The update cycle of the optimization and update module is configurable, with a minimum update cycle of 1 week and a maximum update cycle of 1 month. During the update process, incremental training is used to optimize the SNN model parameters to avoid wasting computing power due to repeated training on historical data. At the same time, the rules in the filtering rule base are dynamically supplemented or adjusted by statistically analyzing the types and characteristics of the initial abnormal data.

[0042] The above is a further description of the present invention in conjunction with specific embodiments, and the scope of protection of the present invention is not limited thereto.

Claims

1. A method for determining abnormal monitoring data in a reservoir, characterized in that, The steps include the following: S1. Construct a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data, and filter out initial abnormal data that do not conform to the rules; S2. Construct a brain-like spiking network (SNN), using the filtered monitoring data time series, the corresponding hydrological parameters, and the instrument operating status parameters as inputs, setting hidden layers, and training the SNN anomaly recognition model with data anomaly classification as the output. S3. Construct a conventional anomaly prediction model to obtain a predicted value as a reference value, introduce "operating condition weight coefficient" and "data credibility weight coefficient", construct W-RMSE, and set the W-RMSE threshold to make a preliminary judgment on data anomalies; S4. The final judgment result is obtained by fusing the SNN anomaly identification result and the W-RMSE preliminary judgment result; S5. Based on the judgment results and manually reviewed data, regularly update the filtering rule base and SNN model parameters.

2. The method for determining abnormal monitoring data of a reservoir according to claim 1, characterized in that, The physical parameter constraints mentioned in step S1 include threshold settings for water level, rainfall, flow rate, and seepage pressure.

3. The method for determining abnormal monitoring data of a reservoir according to claim 1, characterized in that, The inputs in step S2 include water level, rainfall, rainfall level, inflow rate, instrument signal strength, and power supply voltage; the outputs include three types of judgment results: "normal data", "instrument malfunction abnormal data", and "extreme working condition unconventional data".

4. The method for determining abnormal monitoring data of a reservoir according to claim 1, characterized in that, The formula for calculating W-RMSE in step S3 is: ， in: This represents the i-th actual monitoring data. The reference value for the i-th data is generated by an LSTM baseline model trained on historical extreme working condition data from the same period. The working condition weighting coefficient is set based on the intensity of the current extreme working condition. The stronger the working condition, the higher the weighting coefficient. The larger; Data credibility weighting coefficient: set based on the instrument's real-time operating status; n is the number of valid data points after filtering.

5. The method for determining abnormal monitoring data of a reservoir according to claim 1, characterized in that, In step S3, set Threshold is If a certain monitoring data corresponds to If so, it is initially determined to be "normal data"; if If so, it is initially determined to be "suspected abnormal data".

6. The method for determining abnormal monitoring data of a reservoir according to claim 1, characterized in that, The rules for fusion judgment in step S4 are as follows: The confidence threshold for the SNN model classification results is T2. If the probability of the SNN model classifying a certain category is ≥ T2, the result of the SNN model is directly accepted. Otherwise, the classification results are fused according to weights. The classification accuracy and recall of W-RMSE and SNN are calculated through historical sample sets. The weights are assigned using F1 scores, and the two classification results are fused according to the weights.

7. A reservoir anomaly monitoring data determination system, characterized in that, include: The data acquisition module collects monitoring data and instrument operating status data in real time; The filtering rule base construction module is used to build a filtering rule base based on physical parameter constraints and instrument status thresholds for different parameter types of reservoir monitoring data; The initial abnormal data filtering module filters out initial abnormal data that does not conform to the rules; The SNN anomaly detection model construction and anomaly detection module are used to construct a brain-like spiking network (SNN). The filtered monitoring data time series, corresponding hydrological parameters, and instrument operating status parameters are used as inputs. Hidden layers are set, and the SNN anomaly detection model is trained with data anomaly classification as the output, and anomaly detection is performed. Anomaly prediction module: responsible for generating reference values ​​required for W-RMSE calculation. ; The W-RMSE calculation module and anomaly identification module introduce "operating condition weight coefficient" and "data credibility weight coefficient" to calculate W-RMSE and set W-RMSE thresholds for preliminary judgment of data anomalies. The fusion judgment module fuses the SNN anomaly identification results and the W-RMSE preliminary judgment results to obtain the final judgment result; The optimization and update module regularly updates the filtering rule base and SNN model parameters based on the judgment results and manually reviewed data.

8. The reservoir anomaly monitoring data determination system according to claim 7, characterized in that, It also includes a result output and alarm module, which is used to output the final result to the scheduling platform and trigger alarms for abnormal instrument fault data.

9. The reservoir anomaly monitoring data determination system according to claim 7, characterized in that, The aforementioned anomaly prediction module comprises three sub-modules. Data preprocessing submodule: Cleans and normalizes historical extreme operating condition data from the same period; The baseline model training submodule uses a lightweight LSTM network. The input is "the monitoring time series of the past 6 hours + rainfall level + inflow rate", and the output is the predicted value 1 hour later. Real-time prediction submodule: Receives filtered real-time monitoring data, calls the trained baseline model, and outputs reference values ​​for the corresponding time point. The prediction delay is ≤1 minute.

Images