Northwest alpine region adaptive runoff prediction method based on LSTM algorithm

Abstract

The invention discloses a northwest alpine region adaptive runoff prediction method based on an LSTM algorithm, and belongs to the technical field of runoff prediction. The method comprises the following steps: acquiring runoff and influence factor data; screening the key influence factors based on the Spearman correlation coefficient and determining the lagging duration of the key influence factors; constructing an input variable sequence in the comprehensive lagging duration; respectively training a flood season LSTM runoff prediction model and a non-flood season LSTM runoff prediction model; and adaptively selecting a prediction model according to a comparison result of the early-stage runoff volume and a threshold value. According to the method, the problem that a fixed input window and a single model structure are poor in adaptability to high-cold-region runoff prediction in the prior art is solved through differential delay time length analysis and a self-adaptive model selection mechanism, accurate prediction of full-time-period runoff of the northwest high-cold region is achieved, and the flood peak and flood volume prediction precision is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of runoff prediction technology, and particularly relates to an adaptive runoff prediction method based on the LSTM algorithm for the cold northwest region. Background Technology

[0002] In the field of runoff forecasting, data-driven machine learning methods have become an important technological approach. Currently, various machine learning models have been applied to solve hydrological time series forecasting problems. These include traditional machine learning algorithms such as multilayer perceptrons, support vector machines, and decision trees, which achieve runoff forecasting by establishing a nonlinear mapping relationship between hydrological and meteorological input factors and runoff volume. To further address the temporal dependencies in sequence data, long short-term memory (LSTM) neural network models have been introduced into this field. Through gating mechanisms and cell states, these models effectively alleviate the gradient vanishing problem in long-sequence training of traditional recurrent neural networks, enabling them to better capture the temporal characteristics of hydrological sequences and improve the accuracy of runoff forecasting.

[0003] However, existing LSTM-based runoff prediction methods still have significant limitations when applied to the arid and cold regions of Northwest China. First, existing methods typically employ fixed input time windows, failing to fully consider the differential lag effects between various influencing factors and runoff response, resulting in inaccurate temporal correlations between model input features and output targets. Second, existing models are mostly single-model structures, unable to adapt to the significant differences in runoff formation mechanisms between the flood and non-flood seasons in the arid and cold regions of Northwest China, affecting the stability of prediction performance across all time periods. Furthermore, existing methods do not adequately utilize the unique thermal conditions, solar radiation characteristics, and underlying surface factors of the arid and cold regions, limiting the accurate simulation of the complex runoff formation processes in this area. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes an adaptive runoff prediction method for the cold Northwest region based on the LSTM algorithm, thereby resolving the issues present in the existing technologies.

[0005] To achieve the above objectives, this invention provides an adaptive runoff prediction method for the arid northwest region based on the LSTM algorithm, comprising: Acquire time-series runoff data and runoff-related influencing factor data for the target watershed, including meteorological element data, solar radiation characteristic data, and underlying surface condition data; Based on the Spearman correlation coefficient, key influencing factors that are significantly correlated with the runoff time series data are screened from the influencing factor data; Determine the lag duration of each of the key influencing factors, and determine the overall lag duration based on the lag duration of all key influencing factors; A training sample set is constructed using the key influencing factor data and previous runoff data within the comprehensive lag period as input variables and the runoff volume at a predetermined future time as output variable. Based on the training sample set, a flood season LSTM runoff prediction model and a non-flood season LSTM runoff prediction model were trained respectively. Based on the comparison between the current runoff data and the preset runoff threshold, the flood season LSTM runoff prediction model or the non-flood season LSTM runoff prediction model is adaptively selected for runoff prediction.

[0006] Preferably, the underlying surface condition data includes vegetation coverage data, the meteorological element data includes at least one of temperature, humidity and wind speed, and the solar radiation characteristic data includes at least one of sunshine duration and total solar radiation.

[0007] Preferably, the key influencing factors include at least one of the following: average temperature, maximum temperature, minimum temperature, total solar radiation, snow surface temperature, vegetation cover, and anterior runoff.

[0008] Preferably, the step of determining the lag time of each of the key influencing factors includes: For each key influencing factor, calculate the Spearman correlation coefficient between its time series data at different lag days and the runoff time series data; The lag time of the Spearman correlation coefficient falling below the preset correlation threshold for the first time is determined as the lag duration of the key influencing factor.

[0009] Preferably, both the flood season LSTM runoff prediction model and the non-flood season LSTM runoff prediction model are multi-layer LSTM neural network structures, and the LSTM neural network includes at least two hidden layers.

[0010] Preferably, the flood season LSTM runoff prediction model is expressed as follows: In the formula, For the first time during the flood season t Daily runoff forecast for the day, This is a flood season runoff prediction model trained using the LSTM algorithm. For the predicted time t Lag n The first day M Values ​​of influencing factors.

[0011] Preferably, the non-flood season LSTM runoff prediction model is expressed as follows: In the formula, For the non-flood seasont Daily runoff forecast for the day, This is a non-flood season runoff prediction model trained based on the LSTM algorithm.

[0012] Preferably, the step of adaptively selecting the model based on the comparison result between the current runoff data and the preset runoff threshold is as follows: If the runoff volume on day n in the previous runoff data is greater than the preset runoff volume threshold, then the flood season LSTM runoff prediction model is used for prediction. If the runoff volume on day n in the previous runoff data is less than or equal to the preset runoff volume threshold, then the non-flood season LSTM runoff prediction model is used for prediction.

[0013] Preferably, the expression for the adaptively selected model is: In the formula, For the first t Daily runoff forecast for the day, This is an adaptive runoff prediction model trained using the LSTM algorithm, based on the input previous runoff sequence. n The model is selected based on the daily runoff. If the runoff is greater than the preset runoff threshold, the sample is predicted using the flood season runoff prediction model; if the runoff is less than the preset runoff threshold, the sample is predicted using the non-flood season runoff prediction model.

[0014] Preferably, after model prediction, at least one of flood peak error, flood volume error, and peak occurrence time error is calculated to evaluate the model's prediction accuracy for flood elements.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention constructs independent LSTM runoff prediction models for the flood season and non-flood season, and adaptively selects the prediction model based on the comparison results of previous runoff data and preset thresholds. This enables the model to automatically adapt to the runoff formation mechanisms that differ significantly between the flood season and the non-flood season, thereby maintaining high prediction accuracy and stability throughout the entire time period.

[0016] This invention uses Spearman correlation coefficient to determine the lag time between each key influencing factor and the runoff sequence, and constructs an input variable sequence within the comprehensive lag time based on this. This enables the model to more accurately capture the actual delay effect of each influencing factor on runoff, overcomes the limitations of fixed input window, and improves the accuracy of the temporal correlation between model input features and runoff response.

[0017] This invention comprehensively acquires and filters meteorological elements, solar radiation characteristics, and underlying surface conditions as a set of influencing factors to construct training samples. This enables the LSTM model to more fully learn and reflect the unique mixed replenishment mechanism of snowmelt and rainfall in this region, improving the simulation accuracy of complex runoff formation processes and enhancing the model's ability to represent complex underlying surfaces and meteorological conditions in the cold northwest region.

[0018] This invention improves the training speed and convergence stability of the LSTM model by performing linear function normalization on runoff data and influencing factor data, and by using the Adam gradient descent algorithm for model optimization. It also helps to improve the model's predictive performance on unseen data, optimizes model training efficiency, and enhances generalization ability. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram illustrating the material basis and influencing factors of runoff generation in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction of a runoff prediction model for the Northwest Cold Region based on the LSTM algorithm, according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the analysis results of key influencing factors of the Kumarak River runoff in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the change process of the loss function during the training and validation periods of the runoff forecasting model at the Xiehela hydrological station according to an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the predicted results and actual values ​​of the runoff forecasting model test set at the Xiehela Hydrological Station according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 This embodiment provides an adaptive runoff prediction method for the arid northwest region based on the LSTM algorithm, including: Step 1: Obtain runoff time series data and runoff-related influencing factor data for the target watershed. The influencing factor data includes meteorological element data, solar radiation characteristic data, and underlying surface condition data. Furthermore, the underlying surface condition data includes vegetation coverage data, the meteorological element data includes at least one of temperature, humidity and wind speed, and the solar radiation characteristic data includes at least one of sunshine duration and total solar radiation.

[0023] Specifically, in the high-altitude cold regions of Northwest China, runoff is formed by a combination of mountain torrential rains and snowmelt, with snowmelt being the dominant factor. Thermal conditions directly influence the storage, evolution, and physical properties of snow within the controlled catchment area, and are a major factor leading to snowmelt. When temperatures rise, accumulated temperature ≥0℃ appears, and the 0-degree layer height increases rapidly, snowmelt easily forms floods. Sunlight, wind, and humidity are important factors affecting snowmelt. Longer sunshine duration and greater solar radiation provide more heat for snowmelt; stronger winds and faster air movement over the snow and ice surface increase the heat exchange rate, accelerating the melting process; higher humidity leads to faster heat transfer to the ice, thus accelerating the formation of snowmelt-induced floods. Rainfall amount and its spatial and temporal distribution are important driving factors for the rise and fall of flood events. In arid regions, widespread heavy rainfall leads to a rapid increase in surface and groundwater volume. Furthermore, the higher temperature of the rainwater compared to glacial snow helps absorb the potential energy generated by surface condensation, further accelerating snowmelt and resulting in floods. Underlying surface conditions such as watershed morphology, geology, vegetation, and soil conditions influence flood processes from a runoff perspective. Different elevations collect varying amounts of water during rainfall or snowmelt, and soil conditions and vegetation cover also affect evaporation and transpiration rates, altering runoff distribution patterns.

[0024] Based on the river runoff generation process in the high-altitude and cold region of Northwest China, a preliminary set of flood influencing factors was constructed from three aspects: meteorological elements, solar radiation characteristics, and underlying surface conditions. Figure 1 The symbols representing the influencing factors and their data sources are shown in Table 1. When the dynamic changes in the flood discharge capacity of water-related projects are not considered, the underlying surface characteristics within the watershed generally do not have temporal characteristics that change with the flood process. Their impact on runoff prediction results can be regarded as being integrated into the structure and parameters of the machine learning flood forecasting model. Therefore, topographic-related factors are not considered in the runoff prediction model for the time being.

[0025] Table 1

[0026] Step 2: Based on the Spearman correlation coefficient, screen out the key influencing factors that are significantly correlated with the runoff time series data from the influencing factor data; Furthermore, the key influencing factors include at least one of the following: average temperature, maximum temperature, minimum temperature, total solar radiation, snow surface temperature, vegetation cover, and anterior runoff.

[0027] Specifically, the Spearman correlation coefficient is used to measure the strength of the association between the series of influencing factors and the runoff series, extracting key runoff influencing factors and their lagged correlations. A Spearman correlation coefficient greater than 0.8 is considered significant, greater than 0.3 is considered correlated, and less than 0.3 is considered weakly correlated. The calculation formula is as follows:

[0028] In the formula, The Spearman correlation coefficient; and These are the positions of variables X and Y, respectively. and This represents the average rank.

[0029] Step 3: Determine the lag duration of each of the key influencing factors, and determine the overall lag duration based on the lag duration of all key influencing factors; Furthermore, the step of determining the lag time of each of the key influencing factors includes: For each key influencing factor, calculate the Spearman correlation coefficient between its time series data at different lag days and the runoff time series data; The lag time of the Spearman correlation coefficient falling below the preset correlation threshold for the first time is determined as the lag duration of the key influencing factor.

[0030] Lag correlation primarily measures the time delay relationship between runoff and various influencing factors. The runoff series is... The influencing factor sequence is Calculation of runoff and influencing factors with lag k The correlation between the days allows for the construction of new runoff sequences. Construct a new sequence of influencing factors as follows Calculate different lag times k Runoff sequence R 'and influencing factor sequence I The Spearman correlation coefficient is used to search for the critical value of a Spearman correlation coefficient less than 0.3. k cr If the value is given, then the lag correlation between the runoff sequence and this influencing factor is: k cr .

[0031] Step 4: Using the key influencing factor data and previous runoff data within the comprehensive lag period as input variables and the runoff volume at a predetermined future time as output variables, construct a training sample set; Furthermore, before constructing the training sample set, a data preprocessing step is included, which includes normalizing the runoff time series data and the influencing factor data using a linear function normalization method.

[0032] Based on daily runoff data from representative river hydrological stations and meteorological and underlying surface data from the runoff-producing areas, a training, validation, and testing dataset was constructed. The time series was divided into training, validation, and testing periods in an 8:1:1 ratio. Due to the significant difference in magnitude between meteorological and runoff data, a linear function normalization method was used to reduce the impact of this difference on the weight of the runoff input variable. The model output was then inversely normalized before being compared with the observed data.

[0033] In the formula, x 'Indicates normalized data; x Represents observation data; x min and x max These represent the minimum and maximum values ​​in the observed data, respectively.

[0034] LSTM-based runoff forecasting models use samples as input and output data. The samples are combinations of input feature variables and corresponding output target values, determining the parameter optimization process during the model training phase, and thus affecting the reliability and rationality of the simulation and prediction results. The Spearman correlation coefficient between the original runoff and influencing factor sequences is generally >0.3. M As input variables, the lag time of the m-th influencing factor is... k cr,m The overall lag time is Therefore, the model input variables are n Within a day M The sequence of influencing factors is output as the ()th influencing factor sequence. n +1) day's runoff.

[0035] Given the significant differences in runoff during the flood and non-flood seasons in the high-altitude and cold regions of Northwest China, it is necessary to construct separate runoff prediction models for the flood and non-flood seasons. Therefore, it is also necessary to construct separate model samples for the flood season and non-flood season. The criterion for distinguishing between flood season and non-flood season samples is the first... n The daily runoff. Based on river characteristics, determine the... n The daily runoff threshold is Q crIf the first n The day's runoff is greater than Q cr If the sample is included in the flood season model, then the first sample will be used. n The day's runoff is less than Q cr If the sample is not included in the non-flood season model, then the runoff prediction model is used. The runoff prediction model is expressed as:

[0036] In the formula, For the first time during the flood season t Daily runoff forecast for the day, This is a flood season runoff prediction model trained using the LSTM algorithm. For the predicted time t Lag n The first day M Values ​​of influencing factors.

[0037] In the formula, For the non-flood season t Daily runoff forecast for the day, This is a non-flood season runoff prediction model trained based on the LSTM algorithm.

[0038] In the formula, For the first t Daily runoff forecast for the day, For an adaptive runoff prediction model trained using the LSTM algorithm, it is necessary to consider the input previous runoff sequence. n The model is selected based on the daily runoff volume; if it is greater than... Q cr If the sample size is less than 1, then the flood season runoff prediction model should be used for prediction; if the sample size is less than 1, then the sample size should be ... Q cr If the sample is not suitable for flood season runoff prediction, then the non-flood season runoff prediction model should be used for prediction.

[0039] Step 5: Based on the training sample set, train the LSTM runoff prediction model for the flood season and the LSTM runoff prediction model for the non-flood season respectively; Furthermore, the optimization algorithm used when training the LSTM runoff prediction model is the Adam gradient descent algorithm.

[0040] Furthermore, after model training and validation, at least one of the following methods is used to evaluate the model's predictive performance: mean absolute error, mean squared error, and Nash efficiency coefficient.

[0041] Specifically, model hyperparameters mainly include the number of hidden layers, the number of hidden layer neurons, the learning rate, the batch size, the number of iterations, and the optimization algorithm.

[0042] The more hidden layers a model has, the more complex its network becomes, and the stronger its ability to represent complex problems and datasets. The internal connection structure of a multi-hidden-layer LSTM model is as follows: the input layer is connected to the first hidden layer, and the hidden state output by the first hidden layer at each time step... h t As the input to the second hidden layer, the remaining hidden layers are stacked and connected sequentially, finally... H The hidden layer is connected to the output layer.

[0043] The number of hidden layers affects the size of the weights and bias parameters within the flood prediction model; increasing the number of hidden layers leads to a rapid increase in the parameter size. The number of neurons within the hidden layers determines the complexity of the model's internal connection network. The more neurons there are, the larger the size of the connection weight matrix parameters between the input layer, output layer, and hidden layers, and the more complex the model's connection network and corresponding calculations become.

[0044] The learning rate determines the speed at which the model updates its weights during training. Setting the learning rate too high may cause the model to fail to converge, while setting it too low may cause the training process to be too slow. Grid search or random search are commonly used methods to find the optimal learning rate.

[0045] Batch size determines the amount of data input to the model during each training session. While larger batches can improve training speed, they can also lead to insufficient runtime memory. Conversely, smaller batches can improve the model's generalization ability but reduce training speed.

[0046] The number of iterations determines the number of times the model is trained. Too many iterations may lead to overfitting, while too few iterations may lead to underfitting. The appropriate number of iterations can be determined by observing the loss function value on the validation set.

[0047] Regularization is used to control model overfitting, such as L1 and L2 regularization.

[0048] Optimization algorithms are algorithms used to optimize model parameters, such as stochastic gradient descent (SGD) or Adam.

[0049] LSTM models use loss functions to measure model performance and optimization direction, with the commonly used Mean Absolute Error (MAE) being selected. MAE Mean Squared Error MSE ) and Nash efficiency coefficient NSE)The evaluation index is used to assess the prediction results of the LSTM runoff forecasting model, and the calculation formula is as follows:

[0050] In the formula, and These represent the measured runoff sequence and the predicted runoff sequence, respectively. This represents the average of the measured runoff. N Indicates the sequence length. MAE , MSE The smaller the value, the higher the prediction accuracy; NSE The closer the value is to 1, the better the simulation effect.

[0051] To further evaluate the model's prediction accuracy in terms of flood peak, flood volume, and peak time, the following metrics are used: flood peak error, flood volume error, and peak time error.

[0052] In the formula, For the peak flood error, % To measure the peak flow rate, m 3 / s; To predict peak flood flow, m 3 / s.

[0053] In the formula, The maximum 3-day flood volume error is % To measure the maximum three-day flood volume, 100 million cubic meters 3 ; To predict the maximum three-day flood volume, 100 million cubic meters 3 .

[0054] In the formula, The time error for the occurrence of the flood peak is d (days); Measured time of flood peak occurrence, d; Predict the time of the flood peak, d.

[0055] In the formula, The time error for the occurrence of the maximum three-day flood volume is d; The actual maximum three-day flood volume ended in d; Predict the end time of the maximum three-day flood volume, d.

[0056] Flowchart of the runoff prediction model construction based on LSTM algorithm in Northwest China (see) Figure 2 .

[0057] Step 6: Based on the comparison results between the previous runoff data at the current moment and the preset runoff threshold, adaptively select the flood season LSTM runoff prediction model or the non-flood season LSTM runoff prediction model to predict runoff.

[0058] Furthermore, the step of adaptively selecting a model based on the comparison result between the current runoff data and the preset runoff threshold is as follows: If the runoff volume on day n in the previous runoff data is greater than the preset runoff volume threshold, then the flood season LSTM runoff prediction model is used for prediction. If the runoff volume on day n in the previous runoff data is less than or equal to the preset runoff volume threshold, then the non-flood season LSTM runoff prediction model is used for prediction.

[0059] Furthermore, both the flood season LSTM runoff prediction model and the non-flood season LSTM runoff prediction model are multi-layer LSTM neural network structures, and the LSTM neural network includes at least two hidden layers.

[0060] Furthermore, after the model makes its predictions, at least one of the following is calculated: peak error, flood volume error, and peak occurrence time error, in order to evaluate the model's accuracy in predicting flood elements.

[0061] Example 2 This embodiment uses the Kumarak River as an example to apply the method described in Embodiment 1.

[0062] S1. Extraction of key influencing factors; The Kumarak River is the most important tributary of the Aksu River in the northern part of the Tarim River Basin. Its runoff mainly comes from the meltwater of snow and ice from Tomur Peak and Khan Tengri Peak. The Xiehela Hydrological Station is the representative hydrological station. The runoff area of ​​the station is mainly Wensu and Uqturpan counties. Based on the daily runoff data of Xiehela Hydrological Station from January 1, 2000 to May 1, 2023, the daily average temperature, maximum temperature, minimum temperature, average temperature and humidity, precipitation, average wind speed, sunshine duration and total solar radiation of Wensu and Uqturpan county-level meteorological stations, the snow depth, snow water volume and snow surface temperature in the global daily meteorological dataset, and the vegetation coverage data released by the Key Laboratory of Ecological Intelligent Monitoring and Protection of Qinling Mountains in Shaanxi Province of Northwestern Polytechnical University, this study analyzes the key influencing factors of the measured runoff at Xiehela Hydrological Station.

[0063] The Spearman correlation coefficient between the runoff at the Xiehela hydrological station and the collected time series data of influencing factors was calculated. The results of the analysis of key influencing factors of the Kumarak River runoff are shown in [link to relevant data]. Figure 3 Among them, the Spearman coefficients of the runoff sequence and other influencing factors such as previous runoff, average temperature, maximum temperature, minimum temperature, total solar radiation, snow surface temperature, and vegetation cover are all greater than 0.6, and are considered as key influencing factors. Figure 3 The data shows that the lag time for the initial runoff is 25 days, the lag times for the influencing factors at Wensu station are 30 days, 25 days, 35 days, 50 days, 35 days, and 15 days, and the lag times for Wushi station are 30 days, 20 days, 25 days, 45 days, 35 days, and 30 days. The lag time of 50 days is selected as the comprehensive lag time.

[0064] S2. Construction of Kumarak River runoff prediction model; The model architecture was determined to be a four-layer LSTM neural network with two hidden layers and 64 neurons. The input consisted of daily runoff from the Xiehela hydrological station for the previous 50 days, as well as average, maximum, and minimum temperatures, total solar radiation, snow surface temperature, and vegetation cover data for Wensu and Wushi, totaling 650 variables. The output was the runoff for the following day, i.e., one variable. The batch size was 32, the maximum number of training iterations was 100, the learning rate was 0.001, the optimization algorithm was Adam gradient descent, and the activation function was ReLU. The model was divided into training, validation, and testing phases in an 8:1:1 ratio.

[0065] S3, Model Training, Validation, and Testing Effect Analysis.

[0066] The changes in the loss function of the runoff forecasting model at the Xiehela hydrological station during the training and validation periods, such as... Figure 4 As shown, the model converges relatively quickly. A comparison of the predicted results with the actual values ​​on the test set of the runoff forecasting model for the Xiehela hydrological station is shown below. Figure 5 As shown, the prediction process matches the actual process quite well. The runoff prediction model performs well during the prediction period. MAE It is 15.7703m 3 / s, MSE It is 1043.4961. NSE The value is 0.9751, indicating a good overall simulation effect.

[0067] Further analysis of the runoff prediction model's performance in predicting flood peak, peak time, and maximum three-day flood volume showed that the maximum measured flood peak at the Xiehela hydrological station was 1000m. 3 / s, the predicted maximum peak flow is 882.24m³ / s. 3 / s, with an error of 11.78%; the maximum measured three-day flood volume error was 249.26 million m³. 3 The maximum predicted three-day flood volume is 234.18 million cubic meters. 3 The error was 6.05%; the measured and predicted maximum flood peak occurred on July 20, 2022, with an error of 0 days; the measured and predicted maximum three-day flood volume occurred between July 27 and 29, 2022, with an error of 0 days. This shows that the accuracy of the model's predicted maximum three-day flood volume is generally higher than that of the maximum flood peak discharge, and the timing of the flood peak and flood volume is almost error-free.

[0068] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive runoff prediction method based on the LSTM algorithm for the cold northwestern region, characterized in that, Includes the following steps: Acquire time-series runoff data and runoff-related influencing factor data for the target watershed, including meteorological element data, solar radiation characteristic data, and underlying surface condition data; Based on the Spearman correlation coefficient, key influencing factors that are significantly correlated with the runoff time series data are screened from the influencing factor data; Determine the lag duration of each of the key influencing factors, and determine the overall lag duration based on the lag duration of all key influencing factors; A training sample set is constructed using the key influencing factor data and previous runoff data within the comprehensive lag period as input variables and the runoff volume at a predetermined future time as output variable. Based on the training sample set, a flood season LSTM runoff prediction model and a non-flood season LSTM runoff prediction model were trained respectively. Based on the comparison between the current runoff data and the preset runoff threshold, the flood season LSTM runoff prediction model or the non-flood season LSTM runoff prediction model is adaptively selected for runoff prediction.

2. The method according to claim 1, characterized in that, The underlying surface condition data includes vegetation coverage data, the meteorological element data includes at least one of temperature, humidity and wind speed, and the solar radiation characteristic data includes at least one of sunshine duration and total solar radiation.

3. The method according to claim 1, characterized in that, The key influencing factors include at least one of the following: average temperature, maximum temperature, minimum temperature, total solar radiation, snow surface temperature, vegetation cover, and anterior runoff.

4. The method according to claim 1, characterized in that, The step of determining the lag time of each of the key influencing factors includes: For each key influencing factor, calculate the Spearman correlation coefficient between its time series data at different lag days and the runoff time series data; The lag time of the Spearman correlation coefficient falling below the preset correlation threshold for the first time is determined as the lag duration of the key influencing factor.

5. The method according to claim 1, characterized in that, Both the flood season LSTM runoff prediction model and the non-flood season LSTM runoff prediction model are multi-layer LSTM neural network structures, and the LSTM neural network includes at least two hidden layers.

6. The method according to claim 1, characterized in that, The LSTM runoff prediction model for the flood season is expressed as follows: In the formula, For the first time during the flood season t Daily runoff forecast for the day, This is a flood season runoff prediction model trained using the LSTM algorithm. For the predicted time t Lag n The first day M Values ​​of influencing factors.

7. The method according to claim 1, characterized in that, The non-flood season LSTM runoff prediction model is expressed as follows: In the formula, For the non-flood season t Daily runoff forecast for the day, This is a non-flood season runoff prediction model trained based on the LSTM algorithm.

8. The method according to claim 1, characterized in that, The step of adaptively selecting a model based on the comparison between the current runoff data and the preset runoff threshold is as follows: If the runoff volume on day n in the previous runoff data is greater than the preset runoff volume threshold, then the flood season LSTM runoff prediction model is used for prediction. If the runoff volume on day n in the previous runoff data is less than or equal to the preset runoff volume threshold, then the non-flood season LSTM runoff prediction model is used for prediction.

9. The method according to claim 8, characterized in that, The expression for adaptively selecting the model is: In the formula, For the first t Daily runoff forecast for the day, This is an adaptive runoff prediction model trained using the LSTM algorithm, based on the input previous runoff sequence. n The model is selected based on the daily runoff. If the runoff is greater than the preset runoff threshold, the sample is predicted using the flood season runoff prediction model; if the runoff is less than the preset runoff threshold, the sample is predicted using the non-flood season runoff prediction model.

10. The method according to claim 1, characterized in that, After the model makes its predictions, at least one of the following is calculated: peak error, flood volume error, and peak time error, in order to assess the model’s accuracy in predicting flood elements.

Images