A method and device for detecting runoff anomalies at hydrological stations

By combining multi-source feature fusion and physical logic rules, the problem of misjudgment in runoff data anomaly detection in hydropower stations and hydrological stations has been solved, achieving efficient and interpretable anomaly identification and real-time quality control.

CN123115845APending Publication Date: 2026-11-17云南华电金沙江中游水电开发有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611158289.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-31
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for detecting runoff data anomalies in hydropower stations and hydrological monitoring stations suffer from problems such as failing to identify outliers with illogical physical logic and making misjudgments during periods of high fluctuation. Furthermore, deep learning models exhibit vanishing gradients and insufficient accuracy in capturing transient abrupt changes when processing long-term time-series data.

Method used

A multi-source feature fusion method is adopted to reconstruct runoff time series using the TCN-AE model. Anomaly identification is performed by combining MAD dynamic threshold and physical logic rules, including temporal resolution sampling alignment, multi-source physical feature reconstruction, residual dynamic threshold identification, and physical logic rule verification.

Benefits of technology

It enables clear diagnosis of runoff anomalies, possesses physical interpretability and robustness, reduces false alarms and missed alarms, and meets the real-time quality control requirements of hydropower stations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115845A_ABST
    Figure CN123115845A_ABST
Patent Text Reader

Abstract

This invention provides a method and apparatus for detecting measured runoff anomalies at hydrological stations. The method includes the following steps: Step 1: Constructing multi-source features and performing time-resolution sampling alignment and normalization operations to obtain a feature tensor; Step 2: Based on the feature tensor, performing runoff time-series reconstruction based on TCN-AE and multi-source physical features to obtain a reconstructed runoff sequence; Step 3: Based on the reconstructed runoff sequence, performing residual dynamic threshold identification based on MAD to obtain candidate anomalies; Step 4: Based on the candidate anomalies, performing anomaly verification based on physical logic rules to obtain the measured runoff anomaly situation. This invention combines hydrological common sense such as meteorological-water level-flow consistency and runoff generation-confluence causality for posterior arithmetic, so that each detected anomaly point has a clear physical label, greatly improving the decision-making efficiency of hydropower station or hydrological station operation and maintenance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method and device for detecting anomalies in measured runoff at hydrological stations, belonging to the field of hydrological monitoring technology. Background Technology

[0002] Currently, in the actual operation of hydropower stations and hydrological monitoring stations, runoff data is the core input for flood control scheduling, power generation planning, and water resource allocation. However, due to the influence of extreme outdoor environments, sensors are often affected by electromagnetic pulse interference, siltation, obstruction by floating objects, and icing in cold regions, resulting in anomalous values ​​such as sudden spikes, zero-point drift, or false flat-top values.

[0003] In existing technologies, conventional anomaly detection mainly relies on fixed threshold methods for a single variable (such as the 3-Sigma principle). Its shortcomings are: (1) it cannot identify anomalies where the values ​​are within the normal range but the physical logic is inconsistent; (2) in highly fluctuating and non-stationary processes such as flood season, it is easy to judge reasonable values ​​as anomalies. In recent years, with the continuous development of time series prediction and anomaly detection technologies, discriminative models based on deep learning (such as Long Short-Term Memory networks LSTM) have been proposed. However, its recursive structure suffers from the gradient vanishing problem when processing long time series data, and its accuracy in capturing instantaneous change signals is insufficient.

[0004] Therefore, there is an urgent need in this field for an anomaly detection scheme that can integrate physical characteristics such as water level and meteorology, and is physically interpretable and robust. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for detecting anomalies in measured runoff at hydrological stations, which can integrate physical characteristics such as water level and meteorology, and has physical interpretability and improved robustness.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a method for detecting measured runoff anomalies at hydrological stations, comprising the following steps:

[0008] Step 1: Construct multi-source features and perform temporal resolution sampling alignment and normalization operations to obtain the feature tensor;

[0009] Step 2: Based on the feature tensor, perform runoff time series reconstruction based on TCN-AE and multi-source physical features to obtain the reconstructed runoff sequence;

[0010] Step 3: Based on the reconstructed runoff sequence, perform MAD-based residual dynamic threshold identification to obtain candidate outliers;

[0011] Step 4: Based on the candidate anomalies, perform anomaly verification based on physical logic rules to obtain the measured runoff anomaly situation.

[0012] Furthermore, in step 1, multi-source features are constructed and subjected to temporal resolution sampling alignment and normalization operations to obtain a feature tensor, including:

[0013] Acquire multi-source physical time-series data from target hydrological stations in the watershed, the data including at least measured runoff sequences. and station water level sequence Average rainfall sequence of the upstream basin of the monitoring station radar echo sequence of the station and the average temperature sequence of the upstream basin of the monitoring station ;

[0014] Multi-source physical time-series data from different frequencies are sampled to a preset frequency, such as a 1-hour scale. Water level data collected at frequencies higher than the preset frequency (typically, water level data is collected at frequencies higher than 1 hour, such as 15 minutes) are downsampled using the following formula:

[0015] ;

[0016] in and These are water level data at a preset frequency and high-frequency water level data, respectively, where n refers to the number of times the water level data is collected within one cycle of the preset frequency.

[0017] Then, the aligned multidimensional sequences are mapped to the [0,1] interval to construct feature tensors. .

[0018] Furthermore, the temporal resolution sampling alignment and normalization operations also include:

[0019] The time resolution is 1 hour.

[0020] For data with gaps of 3 hours or less, linear interpolation is used; for gaps exceeding 3 hours, the median of the nearest 10-day period is used to fill the gaps.

[0021] Furthermore, in step 2, based on the feature tensor, runoff time series reconstruction based on TCN-AE and multi-source physical features is performed, including:

[0022] feature tensor When input into the TCN-AE model, the reconstructed result is a multivariate sequence whose length is perfectly aligned with the feature tensor in the time dimension. And extract the runoff sequence. ;

[0023] The TCN-AE model includes a symmetrically configured encoder and decoder;

[0024] The encoder utilizes a multi-layer temporal convolutional network (TCN) to convert the input feature tensors into causal convolutions and dilated convolutions. Compressed into a low-dimensional latent space vector,

[0025] The decoder reconstructs the low-dimensional latent space vectors into a sequence consistent with the original input structure, resulting in a multivariate sequence whose length is perfectly aligned with the feature tensor in the time dimension. ;

[0026] The encoder consists of three one-dimensional dilated causal convolutional layers, with a kernel size of [missing information]. expansion factor Set as .

[0027] Each layer contains residual connections and layer normalization.

[0028] To ensure the capture of sufficient past time for the physical characteristics of confluence and generation, its receptive field The calculation formula is:

[0029]

[0030] in, The kernel size is [size]. Let be the i-th expansion factor. Assume the time step is . Hourly (This invention is based on actual traffic flow measurement data, measured once per hour) The actual physical time span of the receptive field. for:

[0031]

[0032] The feature dimension is compressed to the original input feature dimension through a fully connected layer. (generally (Take 4), so that the model retains a representative pattern.

[0033] The structure of the decoder is symmetrical to that of the encoder.

[0034] Furthermore, the TCN-AE model minimizes the reconstruction error between the original input sequence and the reconstructed sequence, allowing the model to utilize the station water level sequence. Average rainfall sequence of the upstream basin of the monitoring station radar echo sequence of the station Average temperature sequence of the upstream basin of the station Reconstructing runoff variation patterns;

[0035] The training data should be taken from at least the past 10 years of actual test data, with the ratio of training set, test set and validation set being 7:1:2.

[0036] Furthermore, in step 3, based on the reconstructed runoff sequence... MAD-based residual dynamic threshold identification is performed to obtain candidate outliers, including:

[0037] First, calculate the residual between the measured runoff and the reconstructed runoff at time t. The formula is as follows:

[0038]

[0039] in The measured runoff at time t, Let be the runoff reconstruction value of TCN-AE at time t. Then, the median absolute deviation (MAD) combined with a sliding window is used. Calculate the dynamic detection threshold :

[0040]

[0041]

[0042] in, Let be the absolute deviation of the median at time t; This represents the median; This is the sensitivity coefficient. If... If it is, then it is determined to be a candidate outlier.

[0043] Furthermore, in step 4, based on the candidate anomalies, anomaly value verification is performed according to physical logic rules to obtain the measured runoff anomaly situation, including:

[0044] Candidate anomalies are input into a preset physical logic gate rule base, and final diagnosis is performed according to the sudden change interference rule, the channel blockage rule, and the zero point drift rule to determine the measured runoff anomaly of the candidate anomalies. The types of measured runoff anomalies include sensor electromagnetic pulse noise sudden change interference anomaly, floating object obstruction or sampling hole blockage anomaly, and sensor hardware temperature drift or zero point offset anomaly.

[0045] Furthermore, the abrupt interference rule includes: If the following discriminant is simultaneously satisfied, the runoff record is confirmed to be abnormal due to abrupt interference from sensor electromagnetic pulse noise:

[0046]

[0047]

[0048]

[0049] in, and These are the water levels (in meters) at time t and time t-1, respectively. Radar echo signal strength ( ).

[0050] Furthermore, the channel blockage rule includes: if the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to obstruction by floating objects or blockage of the sampling orifice:

[0051]

[0052]

[0053]

[0054] in, The variance of runoff over the past four hours ( ), Let t be the rainfall (mm) at time t.

[0055] Furthermore, the zero-point drift rule states that if the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to sensor hardware temperature drift or zero-point offset:

[0056]

[0057]

[0058] in, Let t be the temperature (°C).

[0059] In a second aspect, the present invention provides a device for detecting anomalies in measured runoff at hydrological stations, including a processor and a storage medium;

[0060] The storage medium is used to store instructions;

[0061] The processor is configured to operate according to the instructions to perform the steps of the method according to the first aspect.

[0062] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0063] 1. Diagnostic Clarity: Unlike black-box AI models, this invention combines hydrological common sense such as meteorological-water level-flow consistency and runoff causality for post-test, so that each detected anomaly point has a clear physical label (such as sudden interference, channel blockage, zero drift), which greatly improves the decision-making efficiency of hydropower station or hydrological station operation and maintenance.

[0064] 2. Physical interpretability: This invention effectively solves the pain points of being misjudged as abnormal by traditional threshold methods during special periods and the inability to detect real abnormalities through dual verification by TCN-AE and physical logic gates.

[0065] 3. Strong dynamic adaptability: By introducing a MAD dynamic threshold mechanism based on a sliding window (corresponding to the W-day hydrological cycle), this scheme can automatically adapt to the transition between wet and dry seasons in the watershed. During flood season, the threshold is automatically relaxed to tolerate high-frequency fluctuations in runoff records, and during dry season, the threshold is automatically tightened to capture minute drifts in runoff records, effectively reducing missed and false alarms.

[0066] 4. Computational efficiency: The parallel convolutional characteristics of TCN make the single-point time significantly lower than that of conventional deep learning models such as LSTM, which can meet the real-time requirements of second-level quality control of multiple hydrological monitoring points on the plant side. Attached Figure Description

[0067] Figure 1 This is an overall flowchart of anomaly detection according to an embodiment of the present invention.

[0068] Figure 2 This is the physical logic gate hierarchical review workflow of this invention. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0070] Example 1:

[0071] This embodiment provides a method for detecting measured runoff anomalies at hydrological stations, such as... Figure 1 As shown, it includes the following steps:

[0072] (1) Multi-source feature construction and heterogeneous data processing

[0073] Acquire multi-source physical time-series data from target hydrological stations in the watershed, the data including at least measured runoff sequences. and station water level sequence Average rainfall sequence of the upstream basin of the monitoring station radar echo sequence of the station and the average temperature sequence of the upstream basin of the monitoring station .

[0074] Perform temporal resolution sampling alignment and normalization operations respectively.

[0075] First, data from different frequencies were sampled and standardized, with a time resolution of 1 hour. The high-frequency water level data were then downsampled using the mean, with the following sampling formula:

[0076]

[0077] in and These are water level data at a preset frequency and high-frequency water level data, respectively. Furthermore, linear interpolation is used for data with gaps of 3 hours or less, while data with gaps exceeding 3 hours are filled using the median of the nearest 10 days for the same period.

[0078] Then, the aligned multidimensional sequences are mapped to the [0,1] interval to construct feature tensors. .

[0079] (2) Runoff time series reconstruction based on TCN-AE and multi-source physical characteristics

[0080] feature tensor Input a pre-trained Temporal Convolutional Autoencoder (TCN-AE) model. This model consists of three one-dimensional dilated causal convolutional layers. (The variable is the dependent variable, and the other variables are independent variables) Convolution kernel size expansion factor Set as Each layer contains residual connections and layer normalization. To ensure the capture of runoff generation and confluence physical characteristics over the past M hours (M determined by the time required for runoff confluence in the basin), its receptive field... The calculation formula is:

[0081]

[0082] The dimension is compressed to the input feature dimension through a fully connected layer. (generally (4) is used to filter noise, and then the reconstructed sequence of runoff terms is obtained through a symmetrical decoder structure. .

[0083] (3) Residual dynamic threshold identification based on MAD

[0084] First, calculate the residual between the measured runoff and the reconstructed runoff value at time t. The formula is as follows:

[0085]

[0086] in The measured runoff at time t, Let be the runoff reconstruction value of TCN-AE at time t. Then, the median absolute deviation (MAD) combined with a sliding window is used. Calculate the dynamic detection threshold :

[0087]

[0088]

[0089] in, Let be the absolute deviation of the median at time t; This represents the median; This is the sensitivity coefficient. If... If it is, then it is determined to be a candidate outlier.

[0090] (4) Reviewing outliers based on physical logic rules

[0091] The candidate anomalies are input into the preset physical logic gate rule base, and the final diagnosis is performed according to the sudden interference rule, the channel blockage rule, and the zero drift rule. It is determined whether the anomalies detected in the previous step are the three common anomalies of measured runoff at the station: sensor electromagnetic pulse noise sudden interference anomaly, floating object obstruction or sampling hole blockage anomaly, and sensor hardware temperature drift or zero offset anomaly, or whether they are determined to be non-anomalies.

[0092] 1) Abrupt Interference Rule: If the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to abrupt interference from sensor electromagnetic pulse noise:

[0093]

[0094]

[0095]

[0096] in, and These are the water levels (in meters) at time t and time t-1, respectively. Radar echo signal strength ( The water level showed almost no fluctuation, indicating no actual increase or decrease in river volume; simultaneously, the weak radar echo indicated no strong convective weather over the station, ruling out the possibility of a significant increase in flow due to heavy rainfall. The combination of these two physical factors confirmed that the anomaly in the runoff record was caused by a sudden change in sensor electromagnetic pulse noise.

[0097] 2) Track Blockage Rule: If the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to obstruction by floating objects or blockage of the sampling orifice:

[0098]

[0099]

[0100]

[0101] in, The variance of runoff over the past four hours ( ), The rainfall at time t is in mm. The previously small changes in runoff but large rainfall indicate that the runoff volume should have increased, but this was not detected by the station. The combination of these two physical factors confirms that the abnormal runoff record was caused by obstruction from floating debris or blockage of the sampling orifice.

[0102] 3) Zero-point drift rule: If the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to sensor hardware temperature drift or zero-point offset:

[0103]

[0104]

[0105] in, Let be the air temperature (°C) at time t. Physically, river runoff cannot be negative, but if the river freezes or the water level gauge becomes frozen, the measurement reference may change, resulting in a negative value. A positive reading indicates that water is liquid, and the sensor should be functioning normally. Therefore, the negative runoff value should be due to a baseline shift in the sensor itself caused by environmental temperature changes (temperature drift) or component aging. The combination of these two physical factors confirms that the abnormal runoff record is caused by temperature drift or zero-point offset in the sensor hardware.

[0106] according to Figure 1 The workflow for monitoring measured runoff anomalies, as shown, requires data preprocessing, TCN-AE model training, residual dynamic threshold identification, and physical logic gate hierarchical verification. A specific example is as follows:

[0107] (1) Data acquisition and alignment preprocessing

[0108] This embodiment selects hourly observation data from a hydropower station's inlet monitoring station from 11:00 on July 25, 2005 to 19:00 on October 12, 2025, including runoff, rainfall, water level, radar echo, and temperature. For data with a time resolution higher than 1 hour, downsampling is used for mean normalization. Specifically, for missing values ​​within 3 hours or data with a time resolution lower than 1 hour, linear interpolation is used for completion; for missing values ​​exceeding 3 hours, the median of the same period over the nearest 10 days is used for filling. Finally, Min-Max normalization is used to map all dimensions to the [0,1] space.

[0109] (2) TCN-AE model structure and training settings

[0110] In this embodiment, the TCN encoder employs a three-layer dilated causal convolutional layer. This is achieved by setting the dilation factor. The effective receptive field of the model The calculation is as follows:

[0111]

[0112] This receptive field ensures that the model learns the evolution characteristics of rainfall confluence over the past 15 hours. If the watershed is large and the process from rainfall to confluence to the measuring point takes more than 15 hours, the receptive field can be increased by adding an expansion factor or increasing the number of causal convolutional layers.

[0113] Assuming the time step is Hourly (This invention is based on actual traffic flow measurement data, measured once per hour) The actual physical time span of the receptive field. for:

[0114] 1=15

[0115] The feature dimension is compressed to the original input feature dimension through a fully connected layer. (generally (Take 4), so that the model retains a representative pattern.

[0116] The structure of the decoder is symmetrical to that of the encoder.

[0117] Hourly observation data from the data entry station between 11:00 on July 25, 2005 and 19:00 on October 12, 2025 were manually calibrated and screened to ensure that the remaining data were from normal operating conditions, thus ensuring that the autoencoder learned a physically correct reconstruction. The first 90% of the data was selected as the training set, using the Adam optimizer with a learning rate of 0.001 and a batch size of 64. An early stopping mechanism (Patience=10) was set to prevent overfitting. The loss function was the mean squared error, calculated as follows:

[0118]

[0119] in, It is the loss function.

[0120] (3) Anomaly screening based on MAD threshold

[0121] Sliding window A 168-hour window was selected. The physical basis for choosing this window is that the hydrological sequence at the measuring point is affected by the superposition of natural runoff and upstream hydropower station regulation, as well as by weekly power grid dispatch, resulting in a weekly fluctuation cycle. The threshold was calculated using dynamic MAD over a 168-hour period. It can adaptively adjust within a reasonable range according to the watershed's wet / dry conditions. The value range is [3.0, 4.5]. During the dry season (low mean residual), the threshold is automatically tightened; during the flood season (high mean residual), the threshold is appropriately relaxed.

[0122] (4) Anomaly verification based on physical logic rules

[0123] like Figure 2 As shown, the logic review module diagnoses anomalies based on the following quantitative conditions and reviews them in sequence:

[0124] 1) Physical limit verification (corresponding to zero drift rule): Directly remove measurements that violate thermodynamic logic (such as negative flow rates when the temperature is above 0 degrees).

[0125] 2) "Water level-flow rate" consistency verification (corresponding to sudden change interference rule): Based on the comprehensive system accuracy of 0.02m for the water level gauge, if the flow rate jumps by more than 20% while the water level fluctuates by less than 0.02m and the radar echo signal strength is less than 10dBZ, it is determined that the flow meter pulse is abnormal.

[0126] 3) "Rainfall-flow" consistency verification (corresponding to the channel blockage rule): If there is effective rainfall exceeding 0.1 mm, but the runoff variance is less than 0.01, it is determined that the flow meter is stuck due to floating objects.

[0127] Example 2:

[0128] This embodiment provides a device for detecting abnormal runoff at a hydrological station, including a processor and a storage medium;

[0129] The storage medium is used to store instructions;

[0130] The processor is configured to operate according to the instructions to perform the steps of the method according to Embodiment 1.

[0131] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting measured runoff anomalies at hydrological stations, characterized in that, Includes the following steps: Step 1: Construct multi-source features and perform temporal resolution sampling alignment and normalization operations to obtain the feature tensor; Step 2: Based on the feature tensor, perform runoff time series reconstruction based on TCN-AE and multi-source physical features to obtain the reconstructed runoff sequence; Step 3: Based on the reconstructed runoff sequence, perform MAD-based residual dynamic threshold identification to obtain candidate outliers; Step 4: Based on the candidate anomalies, perform anomaly verification based on physical logic rules to obtain the measured runoff anomaly situation.

2. The method for detecting measured runoff anomalies at hydrological stations according to claim 1, characterized in that, In step 1, multi-source features are constructed and subjected to temporal resolution sampling alignment and normalization operations to obtain the feature tensor, including: Acquire multi-source physical time-series data from target hydrological stations in the watershed, the data including at least measured runoff sequences. and station water level sequence Average rainfall sequence of the upstream basin of the monitoring station radar echo sequence of the station and the average temperature sequence of the upstream basin of the monitoring station ; The sampling of multi-source physical time-series data at different frequencies is unified to a preset frequency, such as a 1-hour scale. Water level data collected at frequencies higher than the preset frequency are downsampled using the mean downsampling formula, as follows: ; in and These are water level data at a preset frequency and high-frequency water level data, respectively, where n refers to the number of times the water level data is collected within one cycle of the preset frequency. Then, the aligned multidimensional sequences are mapped to the [0,1] interval to construct feature tensors. .

3. The method for detecting measured runoff anomalies at hydrological stations according to claim 2, characterized in that, The temporal resolution sampling alignment and normalization operation also includes: The time resolution is 1 hour; For data with gaps of 3 hours or less, linear interpolation is used; for gaps exceeding 3 hours, the median of the nearest 10-day period is used to fill the gaps.

4. The method for detecting measured runoff anomalies at hydrological stations according to claim 1, characterized in that, In step 2, based on the feature tensor, runoff time series reconstruction based on TCN-AE and multi-source physical features is performed, including: feature tensor When input into the TCN-AE model, the reconstructed result is a multivariate sequence whose length is perfectly aligned with the feature tensor in the time dimension. And extract the runoff sequence. ; The TCN-AE model includes a symmetrically configured encoder and decoder; The encoder utilizes a multi-layer temporal convolutional network (TCN) to convert the input feature tensors into causal convolutions and dilated convolutions. Compressed into a low-dimensional latent space vector, The decoder reconstructs the low-dimensional latent space vectors into a sequence consistent with the original input structure, resulting in a multivariate sequence whose length is perfectly aligned with the feature tensor in the time dimension. ; The encoder consists of three one-dimensional dilated causal convolutional layers, with a kernel size of [missing information]. expansion factor Set as Each layer contains residual connections and layer normalization. To ensure the capture of sufficient past time for the physical characteristics of confluence and generation, its receptive field The calculation formula is: ; in, The kernel size is [size]. Let i be the expansion factor; Time step is Hours; the actual physical time span of experiencing the wilderness for: ; The feature dimension is compressed to the original input feature dimension through a fully connected layer. This ensures the model retains a representative pattern; n refers to the number of times water level data is collected within a cycle at a preset frequency. The structure of the decoder is symmetrical to that of the encoder.

5. The method for detecting measured runoff anomalies at hydrological stations according to claim 1, characterized in that, In step 3, based on the reconstructed runoff sequence... MAD-based residual dynamic threshold identification is performed to obtain candidate outliers, including: First, calculate the residual between the measured runoff and the reconstructed runoff at time t. The formula is as follows: ; in The measured runoff at time t, The reconstructed runoff value of TCN-AE at time t; then, the median absolute deviation combined with a sliding window is used. Calculate the dynamic detection threshold : ; ; in, Let be the absolute deviation of the median at time t; This represents the median; This is the sensitivity coefficient; if If it is, then it is determined to be a candidate outlier.

6. The method for detecting measured runoff anomalies at hydrological stations according to claim 5, characterized in that, In step 4, based on the candidate anomalies, anomaly verification is performed according to physical logic rules to obtain the measured runoff anomaly situation, including: Candidate anomalies are input into a preset physical logic gate rule base, and final diagnosis is performed according to the sudden change interference rule, the channel blockage rule, and the zero point drift rule to determine the measured runoff anomaly of the candidate anomalies. The types of measured runoff anomalies include sensor electromagnetic pulse noise sudden change interference anomaly, floating object obstruction or sampling hole blockage anomaly, and sensor hardware temperature drift or zero point offset anomaly.

7. The method for detecting measured runoff anomalies at hydrological stations according to claim 6, characterized in that, The rules for sudden interference include: if the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to sudden interference from sensor electromagnetic pulse noise: ; ; ; in, and These are the water levels at time t and time t-1, respectively. This represents the radar echo signal strength.

8. The method for detecting measured runoff anomalies at hydrological stations according to claim 6, characterized in that, The criteria for track blockage include: if the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to obstruction by floating objects or blockage of the sampling orifice: ; ; ; in, This represents the variance of runoff over the past four hours. Let t be the rainfall at time t.

9. The method for detecting measured runoff anomalies at hydrological stations according to claim 6, characterized in that, Zero-point drift rule: If the following criteria are met simultaneously, the runoff record is confirmed to be abnormal due to sensor hardware temperature drift or zero-point offset: ; ; in, Let t be the temperature at time t.

10. A device for detecting anomalies in measured runoff at a hydrological station, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-9.