Flood event division method and device based on deep learning

Abstract

The disclosure provides a flood episode division method and device based on deep learning, belonging to the technical field of flood data analysis. The method comprises: dividing historical flood data into flood sequence data and non-flood sequence data. The flood sequence data is expanded to obtain target flood sequence data. The target flood sequence data and the non-flood sequence data are spliced according to the time sequence to obtain a flow-rainfall time sequence sample. Based on the flow-rainfall time sequence sample, a multi-layer long short-term memory network is trained using a target loss function to obtain a flood prediction model. The flood prediction model is used to divide the flow-rainfall time sequence data to be divided into flood episodes to obtain a flood episode division result. The flood episode division method and device based on deep learning can improve the accuracy of flood episode identification and division in complex flood scenarios.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the technical field of flood data analysis, and more particularly to a flood event division method and device based on deep learning. BACKGROUND

[0002] Hydrological observation data is usually recorded continuously by year for analysis and statistics. In a year, when a river basin encounters continuous rainfall or superimposed floods from different tributaries, a complex flood process with multiple peaks is often formed. In practical applications, it is necessary to extract the flood events. The extracted flood data can be used to compile hydrological yearbooks, study flood characteristics, and provide samples for parameter calibration of flood forecasting models.

[0003] However, the traditional threshold-based peak screening method is difficult to accurately identify the effective recession process of the peak-to-valley value. The traditional method is prone to miss the "low peak, long duration" flood event, which is of great significance in water resources regulation and ecological impact assessment.

[0004] In summary, the traditional flood event identification and division method cannot adapt to the nonlinear dynamics of flood processes and the time-varying nature of environmental changes, and there is a problem of low accuracy in identifying flood events in complex flood scenarios. SUMMARY

[0005] The purpose of the present disclosure is to provide a flood event division method and device based on deep learning to improve the accuracy of flood event identification and division in complex flood scenarios.

[0006] The first aspect of the embodiment of the present disclosure provides a flood event division method based on deep learning, comprising:

[0007] dividing historical flood data into flood sequence data and non-flood sequence data; data augmentation is performed on the flood sequence data to obtain target flood sequence data; the target flood sequence data and the non-flood sequence data are spliced according to time sequence to obtain flow-rainfall time sequence samples;

[0008] Based on the flow-rainfall time sequence samples, a multi-layer long short-term memory network is trained using a target loss function to obtain a flood prediction model;

[0009] The flood prediction model is used to divide the flow-rainfall time sequence data to be divided into flood events to obtain a flood event division result.

[0010] The second aspect of the embodiment of the present disclosure provides a flood event division device based on deep learning, comprising:

[0011] The data processing module is configured to divide historical flood data into flood sequence data and non-flood sequence data, perform data augmentation on the flood sequence data to obtain target flood sequence data, and splice the target flood sequence data and the non-flood sequence data according to a time sequence to obtain a flow-rainfall time sequence sample.

[0012] The model training module is configured to train a multi-layer long short-term memory network based on the flow-rainfall time sequence sample by using a target loss function to obtain a flood prediction model.

[0013] The model event division module is configured to perform flood event division on to-be-divided flow-rainfall time sequence data by using the flood prediction model to obtain a flood event division result.

[0014] In a third aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the steps of the deep learning-based flood event division method described above when executing the computer program.

[0015] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program implements the steps of the deep learning-based flood event division method described above when executed by a processor.

[0016] The deep learning-based flood event division method and device provided by the embodiments of the present disclosure have the following advantages. On the one hand, the embodiments of the present disclosure model the long-term dependence relationship of the flow-rainfall sequence by using a multi-layer LSTM network for a complex flood scenario, avoid the mechanical division defects of a single threshold, and can improve the flood boundary division accuracy.

[0017] On the other hand, the embodiments of the present disclosure alleviate the model bias caused by scarcity by expanding the flood sample, make the model fully learn the flood peak and base flow transition characteristics, and enhance the generalization ability of the model. The hierarchical feature extraction capability of the multi-layer LSTM can adapt to the hydrological differences of different river basins, accurately divide the complete flood boundary, and improve the accuracy and robustness of model training in combination with the target loss function.

[0018] In summary, the method provided by the embodiments of the present disclosure can improve the accuracy of flood event identification and division in a complex flood scenario, and provide standardized and high-precision results for hydrological yearbook compilation and forecast model calibration. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0020] Figure 1 A flowchart of a flood event division method based on deep learning provided by an embodiment of the present disclosure is shown in FIG. 1.

[0021] Figure 2 A flowchart of a flood event division method based on deep learning provided by another embodiment of the present disclosure is shown in FIG. 2.

[0022] Figure 3 A single-peak and single-event data graph of a specific river basin typical flood event interval provided by an embodiment of the present disclosure is shown in FIG. 3.

[0023] Figure 4 A multi-peak and multi-event data graph of a specific river basin type flood event interval provided by an embodiment of the present disclosure is shown in FIG. 4.

[0024] Figure 5 A deep learning model flood event division result graph of a specific river basin provided by an embodiment of the present disclosure is shown in FIG. 5.

[0025] Figure 6 A structural block diagram of a flood event division device based on deep learning provided by an embodiment of the present disclosure is shown in FIG. 6.

[0026] Figure 7 A schematic block diagram of an electronic device provided by an embodiment of the present disclosure is shown in FIG. 7. DETAILED DESCRIPTION

[0027] In the following description, specific details are set forth in order to provide a thorough understanding of embodiments of the present disclosure. However, persons skilled in the art will understand that the present disclosure can be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure aspects of the present disclosure.

[0028] In order to make the objects, technical solutions, and advantages of the present disclosure clearer, specific embodiments will be described in detail below with reference to the accompanying drawings.

[0029] Reference should be made to Figure 1 , Figure 1 A flowchart of a flood event division method based on deep learning provided by an embodiment of the present disclosure is shown in FIG. 1, which can include S101-S103.

[0030] S101: Divide the historical flood data into flood sequence data and non-flood sequence data. Data augmentation is performed on the flood sequence data to obtain target flood sequence data. The target flood sequence data and the non-flood sequence data are spliced according to the time sequence to obtain the flow-rainfall time sequence sample.

[0031] In this embodiment, the historical flood data is divided into flood sequence data and non-flood sequence data, which includes:

[0032] The time interval corresponding to the historical flood data is divided into a flood episode interval and a non-flood period. The flood episode interval is the time interval in which a flood event occurs. The non-flood period is the time interval in which no flood event occurs.

[0033] The flood data corresponding to the flood episode interval is taken as the flood sequence data.

[0034] The base flow data corresponding to the non-flood period is taken as the non-flood sequence data.

[0035] In this embodiment, before the historical flood data is divided into flood sequence data and non-flood sequence data, the historical flow data and historical rainfall data of the target basin are also preprocessed to obtain the historical flood data.

[0036] In this embodiment, the historical flood data refers to the long-term monitored flow and rainfall time sequence data in the target basin, which contains the original records of flood events and non-flood periods. The target basin refers to a specific basin to be analyzed and referenced. The flood sequence data refers to the flow and rainfall data corresponding to the flood episode interval extracted from the historical data, which contains the complete flood process including rising water, peak value, and falling water. The non-flood sequence data refers to the flow and rainfall data corresponding to the period when no flood event occurs (non-flood period), mainly including base flow or low fluctuation data.

[0037] As shown in the example, Figures 2 to 5 The flood episode division method based on deep learning can include:

[0038] 1. Data collection and data preprocessing:

[0039] For each monitoring station in a specific basin, the time sequence data of its flow and rainfall is collected. Specifically, the historical rainfall data is provided by the rainfall monitoring station, and the corresponding flow data is provided by the flow monitoring station, and the corresponding time point is recorded. The flow data refers to the river flow at each time point, reflecting the flood intensity. The rainfall data refers to the rainfall at each time point, serving as the flood triggering condition.

[0040] The collected data is pre-processed, including removing errors and outliers, such as negative values or unreasonably high values. All data is screened for temporal consistency, and the time stamp format and data frequency are unified. The time stamp refers to a time coordinate accurate to the minute, ensuring temporal consistency of the data.

[0041] Specifically, since the rainfall and flow in natural processes should exhibit continuous and stable changes, sudden changes in values are considered abnormal. The present embodiment can use a moving window method for outlier detection, including:

[0042] (1) Set the window size:

[0043] Select a window size W. For example, W can be the previous 5 time points from the current time point.

[0044] (2) Calculate the sliding window statistics:

[0045] For each time point t, calculate the mean and standard deviation of the data points within the predetermined window W before that time point:

[0046]

[0047]

[0048] wherein is the rainfall or flow data corresponding to time point i.

[0049] (3) Outlier detection:

[0050] Perform outlier detection on the current data point based on the following constraints:

[0051] Condition 1: The current data point suddenly drops to 0, and the previous time point data or the next time point data .

[0052] Condition 2:

[0053] For outliers and missing values in each data sequence, linear interpolation is used for completion. The missing data and data at time point are linearly interpolated using the values , , , , , of the adjacent known data points before and after it.

[0054] The rainfall interpolation formula is:

[0055]

[0056] The flow interpolation formula is:

[0057]

[0058] wherein, is the missing rainfall data at the current time point t, is the previous adjacent rainfall data, is the next adjacent rainfall data, is the missing flow data at the current time point t, is the previous adjacent flow data of t, is the next adjacent flow data of t, is the previous adjacent time point of the missing data, is the next adjacent time point of the missing data.

[0059] (4) Time consistency processing:

[0060] Convert the timestamps of all monitoring sites into a unified format, which is "YYYY-MM-DD HH:MM".

[0061] 2. Flood episode interval division:

[0062] Based on specific research purposes and needs, the flood episode interval is divided for typical single-peak and multi-peak flood events. Each flood event can be represented as a flood episode interval, i.e., the start time and end time of the flood event. Determine the peak flow and peak type characteristics corresponding to each flood episode interval. The peak flow is used to distinguish the flood grade, and the single-peak / multi-peak type, rising and falling water slope, and other peak type characteristics can reflect the flood process morphology. Label the flood episode interval, peak flow, and peak type characteristics to the processed flood data to obtain the flood sequence data. Divide the sequence data outside the flood episode interval into non-flood period to obtain the non-flood sequence data.

[0063] 3. Data augmentation and splicing:

[0064] After obtaining the labeled flood sequence data, the data needs to be augmented to enhance the generalization ability of the model. The data augmentation methods can include stretching and compressing of time series, lengthening and shortening of time length, etc. Splice the augmented data and non-flood period data to obtain complete time series flood division samples.

[0065] For non-flood period data generation, non-flood period data or base flow data with small flow and small fluctuation can be extracted from historical flood data. In the final long-time sample, these non-flood period data are interspersed to enhance the model's ability to identify non-flood periods. To construct longer and more diverse time series samples, the processed multiple types of flood event data and non-flood period data are randomly spliced. The random splicing method can prevent the model from learning fixed flood and non-flood sequence patterns and enhance the model's generalization ability. The spliced data sequence is represented as:

[0066]

[0067] wherein, represents the complete flow-rainfall time series data, containing flood style information, represents random flood event data, represents non-flood period data.

[0068] 4. Data standardization and preparation:

[0069] The flow and rainfall data are standardized to prepare the data for subsequent multilayer long short-term memory network training.

[0070] 5. Construction and training of LSTM network:

[0071] A multilayer LSTM network is constructed and trained using standardized data to capture the time dependence and complex dynamics of flood events. The multilayer LSTM network model architecture includes multiple LSTM layers and Dropout layers to prevent overfitting. Binary cross-entropy is used as the loss function, and the Adam optimizer is used for training. The evaluation metric is binary classification accuracy.

[0072] 6. Model prediction and threshold setting:

[0073] The trained multilayer LSTM model is used to predict new data to obtain the probability of flood event occurrence. By setting an appropriate threshold, the prediction probability is converted into a binary classification result to determine whether a flood event occurs. Finally, a set containing multiple non-overlapping time intervals is obtained. For example, it can be represented as E = {[t1, t2], [t3, t4],..., [ti, tj]}. Each interval [ti, tj] corresponds to a specific flood event, where ti and tj represent the start time and end time of the flood event, respectively.

[0074] S102: Based on the flow-rainfall time series sample, a multilayer long short-term memory network is trained using a target loss function to obtain a flood prediction model.

[0075] In this embodiment, the multilayer long short-term memory network includes:

[0076] the first layer LSTM, the first layer Dropout, the second layer LSTM, the second layer Dropout, the third layer LSTM, the third layer Dropout, and the fully connected layer.

[0077] The first layer LSTM is configured to receive the input time series data and obtain a first sequence output, and the first layer Dropout is configured to randomly drop neurons of the first sequence output to obtain a first input sequence.

[0078] The second layer LSTM is configured to obtain a second sequence output based on the first input sequence, and the second layer Dropout is configured to randomly drop neurons of the second sequence output to obtain a second input sequence.

[0079] The third layer LSTM is configured to obtain and output a hidden state of a last time step based on the second input sequence. The third layer Dropout is configured to randomly drop neurons of the hidden state to obtain a target output.

[0080] The fully connected layer is configured to calculate and output a flood event prediction probability based on the target output.

[0081] In this embodiment, the target loss function is:

[0082]

[0083] wherein, is the target loss function, M is a number of samples contained in a single training batch, is a true label, is a flood event prediction probability.

[0084] In this embodiment, the multi-layer long short-term memory network is a deep learning model based on LSTM units. The multi-layer long short-term memory network enhances the ability to capture long-distance dependencies and complex dynamic characteristics in time series by stacking multiple LSTM layers, and is suitable for time series modeling of flood processes. The target loss function is a function for measuring the difference between the model's predicted flood event probability and the true label, and the target loss function can guide the optimization of model parameters. The input time series data can include flow-rainfall time series samples, flow-rainfall time series data to be divided, and the like.

[0085] For example, the model architecture of the multi-layer long short-term memory network can include:

[0086] The first layer LSTM has an input sequence length of L, a feature dimension of 2, a number of hidden units of 100, and returns an output of the entire sequence.

[0087] For each time step t = 1 to L, the state of the LSTM unit is computed as:

[0088]

[0089] where, represents the input feature vector at the kth time step, represents the hidden state of the previous time step, represents the sigmoid activation function, represents the hyperbolic tangent activation function, represents the element-wise product. W, U, b represent the input weight matrix, recurrent weight matrix, and bias vector of the network, respectively.

[0090] First layer Dropout layer: Add a Dropout layer with a dropout rate of 0.2, which can randomly discard 20% of the neurons to prevent overfitting.

[0091] Second layer LSTM: Similar to the first layer, the number of hidden units is 100, and the output of the entire sequence is returned.

[0092] Second layer Dropout layer: Add a Dropout layer again with a dropout rate of 0.2.

[0093] Third layer LSTM: The number of hidden units is 100, and the output of the entire sequence is not returned, only the hidden state of the last time step is output .

[0094] Third layer Dropout layer: Dropout rate is 0.2.

[0095] Output layer: That is, the fully connected layer, using the sigmoid activation function, outputting the final flood event prediction probability. The flood event prediction probability function is:

[0096]

[0097] where, represents the weight matrix of the output layer, represents the bias vector of the output layer.

[0098] For example, the model compilation process can use the Adam optimizer with a learning rate set to 0.01. Using the training data X and the target result binary label y, train for 200 epochs (epochs) with a batch size of 64 and a validation set ratio of 20%. At each epoch, update the model parameters to minimize the loss.

[0099] S103: Use the flood prediction model to divide the flood event of the to-be-divided flow-rainfall time series data, and obtain the flood event division result.

[0100] In the present embodiment, the flood episode division result is obtained by dividing the to-be-divided flow-rainfall time series data using a flood prediction model, including:

[0101] The flood event prediction probability of the to-be-divided flow-rainfall time series data is calculated using the flood prediction model.

[0102] The target flood event and the target flood episode interval corresponding to the target flood event are determined based on the comparison result of the flood event prediction probability and the episode division threshold.

[0103] All the target flood events and the target flood episode intervals corresponding to the target flood events are taken as the flood episode division result.

[0104] In the present embodiment, the flood prediction model is a multi-layer LSTM network trained by historical flood data, which can capture the time dependence and complex dynamic characteristics of the flow-rainfall time series and output the probability of flood event occurrence. The flow-rainfall time series refers to the target data to be divided into flood episodes, containing time-aligned flow and rainfall sequences, and the format is consistent with the training data. The flood event prediction probability refers to the probability output of the flood prediction model for each time step whether it belongs to a flood event, reflecting the possibility of the time step being in a flood state. The episode division threshold refers to the critical value for converting the prediction probability into a binary classification result.

[0105] The target flood event refers to continuous data with a flood event prediction probability greater than or equal to the episode division threshold, representing a complete flood event. The target flood episode interval refers to the time range corresponding to each target flood event.

[0106] In the present embodiment, the target flood event is determined based on the comparison result of the flood event prediction probability and the episode division threshold, including:

[0107] If the flood event prediction probability is greater than or equal to the episode division threshold, it is determined as a target flood event.

[0108] If the flood event prediction probability is less than the episode division threshold, it is determined as a non-target flood event.

[0109] For example, the trained flood prediction model is used to predict new flow and rainfall data to generate flood event prediction probability. The episode division threshold is set to 0.5, and the flood event prediction probability is converted into a binary classification result. If it is higher than the episode division threshold, it is determined as a target flood event (label 1), and if it is lower than the episode division threshold, it is determined as a non-target flood event (label 0), and finally the occurrence or non-occurrence of the flood event is determined. The flood event discrimination function is:

[0110]

[0111] wherein, represents a label value corresponding to the flood data.

[0112] By acquiring the interval start and end time of each , a set containing multiple non-overlapping time intervals can be obtained, each interval corresponding to a specific flood event.

[0113] From the above, on the one hand, the embodiment utilizes a multi-layer LSTM network to model the long-term dependence relationship of the flow-rainfall sequence for complex flood scenarios, avoiding the mechanical segmentation defects of a single threshold, and can improve the flood boundary division accuracy.

[0114] On the other hand, the embodiment alleviates the model bias caused by scarcity by expanding flood samples, makes the model fully learn the flood peak and base flow transition characteristics, and enhances the generalization ability of the model. The hierarchical feature extraction capability of the multi-layer LSTM can adapt to the hydrological differences of different basins, realize accurate division of the complete flood boundary, and combined with the target loss function, can improve the accuracy and robustness of model training.

[0115] In summary, the embodiment realizes the precision, automation and scenario generalization of flood scene division through the organic combination of data quality optimization, deep model innovation and hydrological logic constraints, and provides efficient and reliable technical support for flood feature analysis, forecast model calibration, etc.

[0116] In an embodiment of the present disclosure, the flood sequence data is data augmented to obtain target flood sequence data, including:

[0117] Based on all flood peak data of the flood sequence data, a first flood peak and a second flood peak are determined.

[0118] Based on the flood sequence data, single-peak flood sequence data is obtained.

[0119] Based on the size relationship of the peak value of the single-peak flood sequence data and the first flood peak and the second flood peak, a single-peak adjustment constraint is determined.

[0120] Based on the single-peak adjustment constraint, a first amplitude scaling factor is generated.

[0121] Based on the first time scaling factor and the first amplitude scaling factor, the single-peak flood sequence data is adjusted, and the adjusted single-peak flood sequence data is added to the target flood sequence data. The first time scaling factor is a preset time adjustment parameter.

[0122] In this embodiment, the first flood peak and the second flood peak are different return period flood peaks determined based on historical flood data through frequency analysis, used to constrain the rationality of data augmentation. For example, the first flood peak can correspond to a 100-year flood peak (return period 100 years, denoted as p = 1%), representing the upper limit of extreme flood intensity. The second flood peak can correspond to a 2-year flood peak (return period 2 years, p = 50%), representing the lower limit of regular flood intensity.

[0123] The single-peak adjustment constraint refers to setting the adjustment range of the first amplitude scaling factor based on the size relationship between the peak of the single-peak flood and the first flood peak and the second flood peak, to ensure that the generated data conforms to the hydrological law. The first time scaling factor is a preset time axis scaling factor used to simulate small changes in flood occurrence speed. The first amplitude scaling factor is a parameter used to adjust the intensity of the single-peak flood sequence data flow, which can generate flood samples of different intensities by small-scale scaling of the original flood flow.

[0124] For example, this embodiment can collect flow observation data in the target basin for many years, and filter out the peak flow of all flood events in history. This embodiment uses the Pearson-III frequency analysis method to statistically fit the historical peak flow, establishes a relationship curve between flow and return period, and extracts the upper limit of extreme intensity of flood peak and the lower limit of regular intensity of flood peak as the first flood peak and the second flood peak, respectively, to dynamically limit the amplitude scaling range and ensure that the flood intensity generated during data augmentation conforms to the natural law of the basin, avoiding unrealistic extreme values.

[0125] This embodiment extracts events containing only a single flood peak from historical flood data, excluding multi-peak flood events. The peak flow of each single-peak flood is calculated, and the single-peak flood is divided into three categories according to its size relationship with the first flood peak and the second flood peak:

[0126] Extreme single-peak flood: peak value exceeds the first flood peak.

[0127] Regular single-peak flood: peak value is between the first flood peak and the second flood peak.

[0128] Low-intensity single-peak flood: peak value is lower than the first flood peak.

[0129] This embodiment sets different single-peak adjustment constraints for different types of single-peak floods. For example, for extreme single-peak floods, the single-peak adjustment constraint is determined as: the interval of the first amplitude scaling factor a is [0.95, 1.05]. For low-intensity single-peak floods, the single-peak adjustment constraint is determined as: the interval of the first amplitude scaling factor is [0.90, 1.10].

[0130] The embodiment presets a first time scaling factor s (s is in the range of [0.95, 1.05]), for example, randomly selects 0.98, 1.02, and calculates an adjusted time axis . The original flow data is interpolated to obtain a new data sequence :

[0131]

[0132] The start time and end time of the flood episode interval are transformed with the first time scaling factor.

[0133] The first amplitude scaling factor is randomly selected within the single-peak adjustment constraint range. The amplitude-adjusted flow data is calculated as follows:

[0134]

[0135] The start time and end time of the flood episode interval remain unchanged because the amplitude scaling does not affect the division of the event interval.

[0136] The embodiment determines different flood peak values of different return periods through frequency analysis, generates reasonable amplitude and time scaling factors accordingly, and adjusts the single-peak flood sequence data. The embodiment can generate diversified flood samples that conform to hydrological laws, alleviate data scarcity, enhance the adaptability of the model to floods of different intensities and durations, and improve the accuracy of flood episode division in complex scenarios.

[0137] In an embodiment of the present disclosure, the flood sequence data is expanded to obtain target flood sequence data, and the method further comprises:

[0138] The multi-peak flood sequence data is obtained based on the flood sequence data.

[0139] The valley points in the multi-peak flood sequence data are extracted, and the multi-peak first adjustment constraint is determined based on the valley points and the base flow threshold.

[0140] The multi-peak second adjustment constraint is determined based on the size relationship between the maximum peak value of the multi-peak flood sequence data and the first and second flood peak values.

[0141] The second amplitude scaling factor is generated based on the multi-peak first adjustment constraint and the multi-peak second adjustment constraint.

[0142] The multi-peak flood sequence data is adjusted based on the second time scaling factor and the second amplitude scaling factor, and the adjusted multi-peak flood sequence data is added to the target flood sequence data. The second time scaling factor is a preset time adjustment parameter.

[0143] In the embodiment, the multi-peak flood sequence data refers to flood event data containing two or more flood peaks, which is manifested as the presence of multiple peaks in the flow hydrograph, and the valley value between peaks does not return to the base flow or there is a significant flow fluctuation. The valley point refers to the lowest point of flow between adjacent flood peaks in a multi-peak flood event, that is, the lowest point of the recession between the two rising processes. The base flow threshold refers to the minimum flow value in the basin during the non-flood period, which is used to determine whether the valley value belongs to an effective recession process. When the valley point is greater than or equal to the base flow threshold, it can be considered as fluctuations of the same flood.

[0144] The multi-peak first adjustment constraint is based on the relationship between the valley point and the base flow threshold, sets a lower limit constraint for the amplitude-scaled valley value, ensures that the adjusted valley value is not lower than the base flow threshold, and maintains the physical rationality of the flood event. The multi-peak second adjustment constraint is a dynamically defined amplitude scaling range according to the size relationship between the maximum peak of the multi-peak flood and the first and second flood peaks, which avoids generating flood samples with super-natural intensity. The second time scaling factor is used to adjust the time axis of the multi-peak flood as a whole, and the second amplitude scaling factor is generated according to the two adjustment constraints and is used to scale the flow intensity of the multi-peak flood in a proportional manner, keeping the relative amplitudes of the peaks unchanged.

[0145] For example, the embodiment can traverse the flood sequence data, extract the valley points (local minimum flow values) between adjacent flood peaks, and record the flow values and time points corresponding to the valley points. Check whether the valley point flow is higher than the base flow threshold. If it is less than the base flow threshold, it is determined as internal fluctuations of the same flood, and the flood is not divided into events.

[0146] For the obtained multi-peak flood sequence data, the multi-peak first adjustment constraint and the multi-peak second adjustment constraint are determined. The multi-peak first adjustment constraint ensures that the data after adjustment needs to be greater than or equal to the base flow threshold, and the lower limit value of the second amplitude scaling factor is obtained. The multi-peak second adjustment constraint sets different ranges of the second amplitude scaling factor according to the size relationship between the maximum peak and the first and second flood peaks. Finally, the range of the second amplitude scaling factor that satisfies the multi-peak first adjustment constraint and the multi-peak second adjustment constraint is determined. The second amplitude scaling factor is generated based on the range.

[0147] The embodiment presets the second time scaling factor to adjust the time axis of the multi-peak flood sequence data, interpolates the adjusted data sequence, and obtains a new multi-peak flood sequence. The second amplitude scaling factor is randomly selected within the range of the multi-peak first adjustment constraint and the multi-peak second adjustment constraint, the flow data after amplitude adjustment of the multi-peak flood sequence data according to the second amplitude scaling factor is calculated, and the expanded multi-peak flood sequence data is obtained.

[0148] The embodiment is aimed at multi-peak flood data, and ensures that the adjusted valley value is reasonable through valley point and base flow threshold constraint, and generates multi-peak samples conforming to hydrological laws by combining the maximum peak value with the relationship between the design flood peak value and the limiting amplitude scaling range.

[0149] The deep learning-based flood event division method corresponding to the above embodiment, Figure 6 The structural block diagram of the deep learning-based flood event division device provided by an embodiment of the present disclosure is shown. For ease of illustration, only parts related to the embodiments of the present disclosure are shown. For reference Figure 6 The deep learning-based flood event division device 20 includes a data processing module 21, a model training module 22, and a model event division module 23.

[0150] The data processing module 21 is configured to divide historical flood data into flood sequence data and non-flood sequence data. The flood sequence data is data augmented to obtain target flood sequence data. The target flood sequence data and the non-flood sequence data are spliced to obtain flow-rainfall time series samples.

[0151] The model training module 22 is configured to train a multi-layer long short-term memory network based on the flow-rainfall time series samples using a target loss function to obtain a flood prediction model.

[0152] The model event division module 23 is configured to divide the flow-rainfall time series data to be divided into flood events using the flood prediction model to obtain a flood event division result.

[0153] In an embodiment of the present disclosure, the data processing module 21 is specifically configured to divide the time interval corresponding to the historical flood data into a flood event interval and a non-flood period. The flood event interval is a time interval in which a flood event occurs. The non-flood period is a time interval in which no flood event occurs.

[0154] The flood data corresponding to the flood event interval is taken as the flood sequence data.

[0155] The base flow data corresponding to the non-flood period is taken as the non-flood sequence data.

[0156] In an embodiment of the present disclosure, the data processing module 21 is specifically further configured to determine a first flood peak value and a second flood peak value based on all flood peak data of the flood sequence data.

[0157] Single-peak flood sequence data is obtained based on the flood sequence data.

[0158] The single-peak adjustment constraint is determined based on the peak value of the single-peak flood sequence data and the size relationship between the first flood peak value and the second flood peak value.

[0159] The first amplitude scaling factor is generated based on the single-peak adjustment constraint.

[0160] The single-peak flood sequence data is adjusted based on the first time scaling factor and the first amplitude scaling factor, and the adjusted single-peak flood sequence data is added to the target flood sequence data. The first time scaling factor is a preset time adjustment parameter.

[0161] In an embodiment of the present disclosure, the data processing module 21 is specifically further configured to obtain multi-peak flood sequence data based on the flood sequence data.

[0162] The valley points in the multi-peak flood sequence data are extracted, and the multi-peak first adjustment constraint is determined based on the valley points and the base flow threshold.

[0163] The multi-peak second adjustment constraint is determined based on the maximum peak value of the multi-peak flood sequence data and the size relationship between the first flood peak value and the second flood peak value.

[0164] The second amplitude scaling factor is generated based on the multi-peak first adjustment constraint and the multi-peak second adjustment constraint.

[0165] The multi-peak flood sequence data is adjusted based on the second time scaling factor and the second amplitude scaling factor, and the adjusted multi-peak flood sequence data is added to the target flood sequence data. The second time scaling factor is a preset time adjustment parameter.

[0166] In an embodiment of the present disclosure, the model training module 22 is specifically configured to include a multi-layer long short-term memory network, which includes:

[0167] The first layer LSTM, the first layer Dropout, the second layer LSTM, the second layer Dropout, the third layer LSTM, the third layer Dropout, and the full connection layer.

[0168] The first layer LSTM is configured to receive input time sequence data and obtain a first sequence output, and the first layer Dropout is configured to randomly discard neurons of the first sequence output to obtain a first input sequence.

[0169] The second layer LSTM is configured to obtain a second sequence output based on the first input sequence, and the second layer Dropout is configured to randomly discard neurons of the second sequence output to obtain a second input sequence.

[0170] The third layer LSTM is configured to obtain and output a hidden state of a last time step based on the second input sequence. The third layer Dropout is configured to randomly discard neurons of the hidden state to obtain a target output.

[0171] The fully connected layer is used to calculate and output the flood event prediction probability based on the target output.

[0172] In one embodiment of this disclosure, the model training module 22 is further configured to have a target loss function as follows:

[0173]

[0174] in, Let M be the target loss function, and M be the number of samples in a single training batch. For real labels, Predicting the probability of flood events.

[0175] In one embodiment of this disclosure, the model field division module 23 is specifically used to calculate the flood event prediction probability of the flow-rainfall time series data to be divided using a flood prediction model.

[0176] The target flood event and the target flood event interval corresponding to the target flood event are determined based on the comparison results between the flood event prediction probability and the event segmentation threshold.

[0177] All target flood events and the corresponding target flood event intervals are used as the flood event segmentation results.

[0178] See Figure 7 , Figure 7 This is a schematic block diagram of an electronic device provided according to an embodiment of the present disclosure. Figure 7 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 6 The functions of the data processing module 21, model training module 22, and model field division module 23 are shown.

[0179] It should be appreciated that the processor 301 in the embodiments of the present disclosure can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0180] The input device 302 can include a touchpad, a fingerprint collection sensor (for collecting fingerprint information and direction information of a fingerprint of a user), a microphone, and the like, and the output device 303 can include a display (LCD, etc.), a speaker, and the like.

[0181] The memory 304 can include a read-only memory and a random access memory, and provide instructions and data for the processor 301. A portion of the memory 304 can also include a non-volatile random access memory. For example, the memory 304 can also store device type information.

[0182] In specific implementations, the processor 301, the input device 302, and the output device 303 described in the embodiments of the present disclosure can execute the implementation manners described in the first embodiment and the second embodiment of the flood event division method based on deep learning provided by the embodiments of the present disclosure, and can also execute the implementation manners of the electronic device 300 described in the embodiments of the present disclosure, which will not be described here.

[0183] In another embodiment of the present disclosure, a computer readable storage medium is provided, which stores a computer program. The computer program includes program instructions, which, when executed by a processor, implement all or part of the processes of the above-mentioned embodiment methods. The computer program can also instruct related hardware to complete the above-mentioned processes. The computer program can be stored in a computer readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0184] The computer readable storage medium can be an internal storage unit of the electronic device, such as a hard disk or a memory of the electronic device. The computer readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of the electronic device. The computer readable storage medium is used to store the computer program and other programs and data required by the electronic device. The computer readable storage medium can also be used to temporarily store data that has been output or will be output.

[0185] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present disclosure.

[0186] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the electronic device and the units described above can refer to the corresponding processes in the above-mentioned method embodiments, which will not be described here.

[0187] In several embodiments provided in the present application, it should be understood that the disclosed electronic device and method can be implemented in other manners. For example, the embodiments of the apparatus described above are merely illustrative. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between the units can be indirect coupling or communication connection through some interfaces, or can be in electrical, mechanical or other forms.

[0188] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present disclosure.

[0189] In addition, each functional unit in the various embodiments of the present disclosure can be integrated in one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware, or in the form of a software functional unit.

[0190] The above is merely specific embodiments of the present disclosure, but the protection scope of the present disclosure is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present disclosure, and these modifications or replacements should be covered in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A method for flood event division based on deep learning, characterized in that, The method comprises the following steps: dividing historical flood data into flood sequence data and non-flood sequence data; data augmentation is performed on the flood sequence data to obtain target flood sequence data; the target flood sequence data and the non-flood sequence data are spliced according to time sequence to obtain flow-rainfall time sequence samples; a flood prediction model is trained by using a target loss function based on the flow-rainfall time sequence samples and a multi-layer long short-term memory network, wherein the multi-layer long short-term memory network comprises a first layer LSTM, a first layer Dropout, a second layer LSTM, a second layer Dropout, a third layer LSTM, a third layer Dropout and a full connection layer; the flood prediction model is used to divide the flow-rainfall time sequence data to be divided into flood stages to obtain a flood stage division result; the data augmentation on the flood sequence data to obtain the target flood sequence data comprises the following steps: determining a first flood peak value and a second flood peak value based on all flood peak data of the flood sequence data; obtaining single-peak flood sequence data based on the flood sequence data; determining a single-peak adjustment constraint based on the peak value of the single-peak flood sequence data and the size relationship between the first flood peak value and the second flood peak value; generating a first amplitude scaling factor based on the single-peak adjustment constraint; adjusting the single-peak flood sequence data based on a first time scaling factor and the first amplitude scaling factor, and adding the adjusted single-peak flood sequence data to the target flood sequence data; the first time scaling factor is a preset time adjustment parameter; obtaining multi-peak flood sequence data based on the flood sequence data; extracting valley points in the multi-peak flood sequence data, and determining a multi-peak first adjustment constraint based on the valley points and a base flow threshold value; determining a multi-peak second adjustment constraint based on the maximum peak value of the multi-peak flood sequence data and the size relationship between the first flood peak value and the second flood peak value; generating a second amplitude scaling factor based on the multi-peak first adjustment constraint and the multi-peak second adjustment constraint; adjusting the multi-peak flood sequence data based on a second time scaling factor and the second amplitude scaling factor, and adding the adjusted multi-peak flood sequence data to the target flood sequence data; the second time scaling factor is a preset time adjustment parameter. 2.The deep learning-based flood event partitioning method of claim 1, wherein, The historical flood data is divided into flood sequence data and non-flood sequence data, which comprises the following steps: dividing the time interval corresponding to the historical flood data into flood stage intervals and non-flood periods; the flood stage interval is a time interval in which a flood event occurs; the non-flood period is a time interval in which no flood event occurs; the flood data corresponding to the flood stage interval is taken as the flood sequence data; the base flow data corresponding to the non-flood period is taken as the non-flood sequence data. 3.The deep learning based flood event partitioning method of claim 1, wherein, The first layer LSTM is used to receive input time sequence data and obtain first sequence output, and the first layer Dropout is used to randomly discard neurons of the first sequence output to obtain first input sequence; The second layer LSTM is configured to obtain a second sequence output based on the first input sequence, and the second layer Dropout is configured to randomly drop neurons of the second sequence output to obtain the second input sequence; The third layer LSTM is configured to obtain and output a hidden state of a last time step based on the second input sequence; The third layer Dropout is configured to randomly drop neurons of the hidden state to obtain a target output; The full connection layer is configured to calculate and output a flood event prediction probability based on the target output. 4.The deep learning based flood event partitioning method of claim 1, wherein, The target loss function is: wherein, is the target loss function, M is the number of samples in a single training batch, is the true label, is the flood event prediction probability. 5.The deep learning based flood event partitioning method of claim 1, wherein, The flood prediction model is used to divide the to-be-divided flow-rainfall time series data into flood episode division results, including: The flood prediction model is used to calculate a flood event prediction probability of the to-be-divided flow-rainfall time series data; Based on a comparison result of the flood event prediction probability and a flood episode division threshold, a target flood event and a target flood episode interval corresponding to the target flood event are determined; All target flood events and target flood episode intervals corresponding to the target flood events are taken as the flood episode division results.

6. A flood event division device based on deep learning, characterized in that, The data processing module is configured to divide historical flood data into flood sequence data and non-flood sequence data; The flood sequence data is subjected to data augmentation to obtain target flood sequence data; and the target flood sequence data and the non-flood sequence data are spliced according to a time sequence to obtain flow-rainfall time series samples. The data processing module is specifically configured to determine a first flood peak value and a second flood peak value based on all flood peak values of the flood sequence data; Single-peak flood sequence data is obtained based on the flood sequence data; A single-peak adjustment constraint is determined based on a peak value of the single-peak flood sequence data and a size relationship between the first flood peak value and the second flood peak value; A first amplitude scaling factor is generated based on the single-peak adjustment constraint; The single-peak flood sequence data is adjusted based on a first time scaling factor and the first amplitude scaling factor, and the adjusted single-peak flood sequence data is added to the target flood sequence data; The first time scaling factor is a preset time adjustment parameter; Multi-peak flood sequence data is obtained based on the flood sequence data; Valley points in the multi-peak flood sequence data are extracted, and a multi-peak first adjustment constraint is determined based on the valley points and a base flow threshold; A multi-peak second adjustment constraint is determined based on a maximum peak value of the multi-peak flood sequence data and a size relationship between the first flood peak value and the second flood peak value; A second amplitude scaling factor is generated based on the multi-peak first adjustment constraint and the multi-peak second adjustment constraint; The multi-peak flood sequence data is adjusted based on a second time scaling factor and the second amplitude scaling factor, and the adjusted multi-peak flood sequence data is added to the target flood sequence data; The second time scaling factor is a preset time adjustment parameter; ​ The model training module is configured to train a multi-layer long short-term memory network based on the flow-rainfall time series sample and by using a target loss function, so as to obtain a flood prediction model; the multi-layer long short-term memory network comprises a first layer LSTM, a first layer Dropout, a second layer LSTM, a second layer Dropout, a third layer LSTM, a third layer Dropout, and a full connection layer; The model event division module is configured to perform flood event division on the flow-rainfall time series data to be divided by using the flood prediction model, so as to obtain a flood event division result.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the steps of the method according to any one of claims 1 to 5.

Images