Substation waterlogging space-time prediction method based on machine learning

Through machine learning methods, combined with random forests and Informer models, a space-time prediction system for substation flooding is built, which solves the problems of complex calculations and insufficient timeliness of traditional numerical models, and realizes accurate prediction of flood location and depth and personalized flood prevention strategies, which improves flood prevention ability.

CN119026934APending Publication Date: 2024-11-26CHINA THREE GORGES UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410930289.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The traditional numerical model is complex and time-consuming to calculate in substation floodwater prediction, which cannot meet the timeliness of flood control and neglects the uncontrollability of the floodwater occurrence area and lacks a large-scale spatio-temporal prediction model.

Method used

Using machine learning-based methods, targeted flood control treatment measures are formulated through data collection and preprocessing, construction of floodwater coupling models, designing rainfall plans, building flood space and timing prediction models, model combination and optimization, model evaluation and application, and targeted flood control treatment measures are formulated and used. Random forest and Informer models implement spatiotemporal prediction.

Benefits of technology

Accurate prediction of the changes in the location and depth of substation flooding over time is achieved, comprehensive space-time prediction information is provided, and the waterlogging risk assessment and flood prevention support capabilities are improved, and early warning and personalized flood prevention strategies can be provided in flooding incidents.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026934A_ABST
    Figure CN119026934A_ABST
Patent Text Reader

Abstract

The invention discloses a transformer substation waterlogging space-time prediction method based on machine learning. The method comprises the following steps: S1, data collection and preprocessing; s2, constructing a transformer substation waterlogging coupling model; s3, designing different rainfall schemes to generate rainfall-waterlogging samples; s4, constructing a waterlogging space prediction model; s5, constructing a waterlogging time sequence prediction model; s6, model combination and optimization; s7, performing model evaluation and application; s8, formulating targeted flood prevention treatment measures; through the machine learning technology, the combined model RF-Informer model is built, the change of the inland inundation position and the inland inundation depth of the transformer substation along with time can be predicted, comprehensive space-time prediction information is provided, and the comprehensive assessment of the inland inundation risk is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of substation waterlogging prediction, and in particular to a substation waterlogging spatiotemporal prediction method based on machine learning. Background Art

[0002] In recent years, with the intensification of climate warming, the global water cycle has accelerated, and the meteorological disasters caused by frequent extreme weather and climate events have continued to intensify, seriously threatening the normal and stable operation of the power grid. As an important part of the power system, substations undertake the key tasks of power conversion, transmission and distribution. However, substations are often affected by natural disasters such as waterlogging, which brings serious safety hazards and economic losses to the operation of the power grid. Waterlogging may not only cause damage to substation equipment, but also affect the reliability of power supply, causing inconvenience to the production and life of nearby residents and enterprises.

[0003] At present, waterlogging risk analysis mainly uses numerical models, including hydrological models, hydrodynamic models, and hydrological and hydrodynamic models. These models have their own advantages and disadvantages: hydrological models have high computational efficiency, but cannot calculate the evolution of floods; hydrodynamic models can calculate hydraulic elements such as flow velocity and water depth within the grid, but their computational efficiency is low; hydrological and hydrodynamic models use hydrological models coupled with two-dimensional hydrodynamic models to ensure model accuracy and good computational efficiency, and are a hot topic in flood model research.

[0004] However, traditional numerical models still have limitations in waterlogging prediction. They often require a large amount of meteorological data, hydrological data and topographic data as input, and complex mathematical equations to solve. The calculation process is complicated and time-consuming, and cannot meet the timeliness requirements of substation flood emergency response.

[0005] In recent years, in order to make up for the shortcomings of numerical models, scholars have gradually applied machine learning methods to the prediction of urban waterlogging disasters. At present, the prediction of urban waterlogging depth is mainly aimed at a single flood-prone point, but to a certain extent ignores the uncontrollability of urban waterlogging occurrence areas at different times, and easily ignores the urban waterlogging situation in other areas. There is a lack of large-scale urban waterlogging spatiotemporal prediction models. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a spatiotemporal prediction method for substation waterlogging based on machine learning to solve the problems of the background technology.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for spatiotemporal prediction of substation waterlogging based on machine learning, which comprises the following steps:

[0008] S1, data collection and preprocessing;

[0009] S2, construct substation waterlogging coupling model;

[0010] S3, design different rainfall schemes to generate rainfall-waterlogging samples;

[0011] S4, construct a spatial prediction model for waterlogging;

[0012] S5. Construct a time series prediction model for waterlogging;

[0013] S6, model combination and optimization;

[0014] S7, Model evaluation and application;

[0015] S8. Develop targeted flood prevention and treatment measures.

[0016] Preferably, the S1 specifically includes:

[0017] 1.1、Data Collection:

[0018] First, the surface elevation data of the substation area is scanned to provide detailed terrain information of the substation and its surrounding areas to ensure accurate simulation of water flow paths and waterlogging locations. Next, detailed information on the layout and flow of the drainage network in the substation area is collected to understand the capacity and efficiency of the drainage system and provide basic data for simulating drainage and waterlogging. In addition, the spatial distribution, structural characteristics and usage data of roads and buildings around the substation are collected to identify obstacles that may affect water flow and waterlogging, providing accurate spatial information for the hydrodynamic model. Finally, boundary condition data, including meteorological data and hydrological data, are collected to provide external driving conditions for the hydrodynamic model and ensure the authenticity and reliability of the simulation results.

[0019] 1.2 Data processing and integration:

[0020] In the data cleaning stage, statistical analysis and graphic visualization methods are used to detect and remove noise and error values, and to fill in missing data. For surface elevation data, abnormal elevation points are removed, and interpolation methods are used to fill in missing elevation data. For drainage network data, incorrect pipe diameter and material information are corrected to ensure the consistency of flow data.

[0021] During the format conversion process, data from different sources are converted into a unified file format; the original LiDAR point cloud data is converted into a digital elevation model, and the drainage network map in PDF format is converted into GIS data in vector format. This process ensures the consistency between different data sets and facilitates data integration and processing;

[0022] Finally, all data are converted to the unified WGS84 coordinate system to ensure spatial matching between different data sets; the surface elevation data is converted from the local coordinate system to the WGS84 coordinate system, and the drainage network layout data is aligned to the unified coordinate system to ensure consistency with other spatial data.

[0023] Preferably, the S2 specifically includes:

[0024] First, the model was constructed using MIKE 21 software; necessary input data were prepared, including surface elevation data, drainage network information, spatial distribution data of roads and buildings, and boundary condition data;

[0025] Secondly, the mathematical description of the model is based on the water flow continuity equation, the water flow motion equation and the bed friction slope, and the formula is as follows:

[0026] Continuity equation for water flow:

[0027]

[0028] Where h is the water depth, u is the depth-averaged velocity along the x direction, v is the depth-averaged velocity along the y direction, t is time, and x and y are spatial coordinates;

[0029] The equation of motion for water flow is:

[0030] The momentum equation in the x direction is:

[0031]

[0032] The momentum equation in the y direction is:

[0033]

[0034] Where g is the acceleration due to gravity, S bx and S by are the components of the bed slope in the x and y directions, S fx and S fy are the components of the bed surface friction slope in the x and y directions, respectively;

[0035] Bed surface friction slope:

[0036] The bed surface friction slope is usually calculated using the Manning formula:

[0037]

[0038] Where n is the Manning roughness coefficient;

[0039] Parameter calibration process: Parameter calibration is to adjust the internal parameters of the model by comparing with the actual observation data, so that the simulation results are as consistent as possible with the measured data; this process usually relies on historical rainfall events and determines the best parameter combination by minimizing the difference between the model prediction value and the measured value; the method used is to minimize the residual sum of squares:

[0040]

[0041] Among them, M i is the model prediction value, O i is the corresponding measured value; this process involves repeatedly adjusting the model parameters until the model's prediction results reach satisfactory consistency and accuracy with the measured data;

[0042] Conduct model validation: The validation process tests the predictive ability of the parameterized model by using independent rainfall events. The purpose of the validation is to confirm whether the model can stably predict the hydrodynamic response in the substation area under different rainfall conditions. The validation results can evaluate the reliability and applicability of the model.

[0043] Through the above steps, a two-dimensional hydrodynamic model based on MIKE 21 was constructed, and its accuracy and reliability in simulating the hydrodynamic response under different rainfall conditions were ensured through parameter calibration and verification.

[0044] Preferably, the S3 specifically includes:

[0045] 3.1 Rainfall scheme design:

[0046] Select rainfall type and characteristics: Select appropriate rainfall type according to the climate and hydrological characteristics of the region;

[0047] Setting rainfall intensity and duration: Setting the rainfall intensity R0 and duration τ. These parameters drive the rainfall process through mathematical models and affect the flow and accumulation of surface water.

[0048] The exponential decay model of rainfall intensity is expressed as:

[0049]

[0050] Among them, R(t) is the rainfall intensity at time t, R0 is the initial rainfall intensity, and τ is the time constant, which describes the decay rate of rainfall intensity over time;

[0051] Spatial distribution characteristics: Considering the spatial distribution characteristics of rainfall in the study area, different distribution patterns will have a significant impact on surface runoff and waterlogging risks;

[0052] 3.2 Simulate and generate rainfall-waterlogging samples:

[0053] The designed rainfall scheme was input into the MIKE 21 hydrodynamic model for numerical simulation to obtain rainfall-waterlogging sample data under different rainfall conditions, which served as the basis for subsequent model training and verification.

[0054] Preferably, the S4 specifically includes:

[0055] 4.1 Dataset Construction:

[0056] First, the grid waterlogging spatial feature data and simulation result data are obtained to form a training data set; the grid waterlogging spatial feature data includes terrain, drainage network, road and building information, while the simulation result data comes from the simulation results of the MIKE 21 model under different rainfall conditions. These data are combined to form a data set, which is preprocessed, including data cleaning, format conversion, feature extraction and normalization steps; the preprocessed data set is divided into a training set (80%) and a validation set (20%) for model training and validation;

[0057] Normalization formula:

[0058]

[0059] The normalized data range is [0, 1], where min(x) and max(x) are the minimum and maximum values ​​of the sample data, respectively;

[0060] 4.2 Establishment of spatial prediction model for waterlogging:

[0061] A spatial prediction model for substation waterlogging was established based on the random forest algorithm;

[0062] Decision tree construction:

[0063] For each decision tree T i , the training data set is randomly sampled to obtain a subset D i , randomly select a subset F from the feature set F i Conduct training; the process of building a decision tree is as follows:

[0064] T i =BuildTree(D i , F i )

[0065] Among them, BuildTree is the decision tree construction algorithm, D i is a sample subset, F i is a feature subset;

[0066] A collection of trees:

[0067] A random forest contains N decision trees, forming a collection of trees:

[0068] RF={T1,T2,...,T N}

[0069] predict:

[0070] For a new input feature vector X, each tree T i Give a prediction result P i , the prediction results of all trees are obtained by majority voting or averaging to get the final result P:

[0071]

[0072] Among them, P is the probability of waterlogging, which ranges from 0 to 1. The closer it is to 1, the greater the possibility of waterlogging. X is the input feature vector, including terrain, drainage facilities, buildings, and road features.

[0073] During the model training process, the preprocessed training data set was used to train the random forest model to predict the probability of waterlogging in different areas of the substation;

[0074] 4.3 Results Visualization:

[0075] In order to more intuitively display the possibility of waterlogging, color mapping technology is used to visualize the prediction results.

[0076] Preferably, the S5 specifically includes:

[0077] 5.1 Dataset Construction:

[0078] Firstly, the prediction results of waterlogging locations are obtained from the waterlogging spatial prediction model, and combined with the simulation results of the corresponding grid waterlogging points, a waterlogging time series dataset is generated and preprocessed, including data cleaning, format conversion, feature extraction and normalization steps; the preprocessed dataset is divided into a training set (80%) and a validation set (20%) for model training and validation;

[0079] 5.2 Establishment of waterlogging time series prediction model:

[0080] In order to build a waterlogging time series prediction model, the Informer model is used. The design goal of the Informer model is to improve the computational efficiency and accuracy of long time series prediction through self-attention mechanism, multi-layer network stacking and step-by-step decoding method.

[0081] The Informer model includes an encoder and a decoder. The encoder receives a large amount of long sequence input and uses a multi-head sparse self-attention mechanism to replace the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and fills the target elements with 0 values.

[0082] The encoder consists of multiple multi-head sparse self-attention and distillation; the calculation formula of the multi-head sparse self-attention mechanism is:

[0083]

[0084] Where: Q, K and V are matrices obtained by different linear transformations of the input variables; is the matrix obtained after Q is probabilistically sparse; d is the dimension of the query vector and the key-value vector; Softmax is the activation function;

[0085] The multi-head sparse self-attention is composed of several stacked sparse self-attention layers; there are multiple multi-head sparse self-attention layers in the encoder; when the input data is transferred from the i-th multi-head sparse self-attention layer to the i+1-th multi-head sparse self-attention layer, a distillation operation is performed; the purpose of the distillation operation is to reduce the complexity of the model and reduce the amount of calculation, giving higher weights to the main features, and the process of extracting features from the j-th layer to the j+1-th layer will be simplified as follows to speed up the operation; after encoding the input sequence, the encoder generates a concatenated feature map and sends it to the multi-head self-attention layer of the decoder;

[0086]

[0087] Where: [*] is the attention module in multi-head sparse self-attention; Conv1d represents the regularized convolution operation on the time series; ELU is the activation function; MaxPool represents the maximum pooling operation;

[0088] The "distillation" operation adds a pooling layer with a stride of 2, making Downsample to 1 / 2, thereby reducing memory utilization to O((2–λ)LlogL), where the value of λ is very small; to enhance the robustness of the distillation operation, the Informer model constructs one copy by halving the input, and reduces the number of attention blocks to align the output dimensions; the decoder part includes Masked multi-head sparse self-attention blocks and multi-head attention blocks. The attention mechanism is to prevent each position from paying attention to future information and avoid autoregression; the Masked multi-head sparse self-attention block receives the label vector of the following formula, and the decoder uses a generative method for decoding; the decoder uses the intermediate result output by the encoder to perform a multi-head attention operation, adjusts the dimension of the output data through the fully connected layer, and finally outputs the result of the predicted sequence length at one time;

[0089]

[0090] Where: is the t-th input sequence of the decoder; is the start mark of the tth sequence; is the placeholder for the target sequence of the tth sequence;

[0091] According to the location of waterlogging predicted by the random forest algorithm, a waterlogging time series data set is generated in combination with the corresponding grid simulation results. This data set includes the changes in the waterlogging depth in different areas over time, that is, the data on the waterlogging depth in different areas of the substation at different times; a waterlogging time prediction model is established based on the Informer time series model, and the generated waterlogging time series data is used as input to predict the changing trend of the waterlogging depth in different areas of the substation over time, so as to take preventive measures and respond to waterlogging risks in a timely manner.

[0092] Preferably, the S6 specifically includes:

[0093] 6.1 Model Combination:

[0094] The spatial waterlogging prediction model based on the Random Forest (RF) algorithm is combined with the temporal waterlogging prediction model based on the Informer time series model. The combined model is optimized using a sequential concatenation method to improve prediction accuracy and practicability.

[0095] First, the RF model is used to predict the spatial distribution of waterlogging; the RF model constructs multiple decision trees and makes predictions through a majority voting mechanism;

[0096] Secondly, the waterlogging locations predicted by the RF model are combined with the simulation results of the corresponding grid waterlogging points to generate a waterlogging time series dataset. This dataset is input into the Informer model to predict the change of waterlogging depth over time in different areas of the substation.

[0097] Through this sequential cascade method, the high-risk areas of waterlogging are first determined by the RF model, and then the changing trend of waterlogging depth in these areas over time is predicted by the Informer model, thus achieving more accurate spatiotemporal waterlogging prediction;

[0098] 6.2 Optimization Model:

[0099] Parameter tuning of random forest (RF) model:

[0100] The random forest model is an integrated model composed of multiple decision trees. Its performance is affected by multiple hyperparameters. The hyperparameters and their tuning methods are as follows:

[0101] Number of trees: The more trees there are, the more stable the model is, but the computational cost is also higher; use grid search or random search to find the optimal value within a certain range;

[0102] Maximum depth of the tree: Control the maximum depth of each tree to prevent overfitting; deeper trees can capture complex patterns but are prone to overfitting; select the optimal depth through cross-validation;

[0103] Parameter tuning of the Informer model:

[0104] The hyperparameters and their tuning methods are as follows:

[0105] The number of encoder and decoder layers: The more layers there are, the stronger the model's expressiveness is, but it is prone to overfitting; select the optimal number of layers through cross-validation;

[0106] Number of attention heads: The number of heads in the multi-head attention mechanism. Increasing the number of heads can capture more features, but the amount of computation also increases. The optimal number of heads is selected through experiments.

[0107] Preferably, the S7 specifically includes:

[0108] The validation set data was used to evaluate the model under the optimal parameters, and the prediction accuracy of the model was evaluated by comparing it with the actual observed data. The mean square error (MSE), mean absolute error (MAE), and goodness of fit R 2 As a model evaluation indicator;

[0109]

[0110] Where n is the number of prediction points; y i are the predicted value and true value of the i-th point data in the test data set respectively; is the mean of the observed values;

[0111] After the evaluation is completed, the optimal combined model will be applied to the actual substation waterlogging prediction to provide accurate support and decision-making basis for substation waterlogging management and disaster prevention and mitigation.

[0112] Preferably, the S8 specifically includes:

[0113] According to the characteristics of different areas of the substation and the different requirements for waterlogging, formulate corresponding flood prevention and treatment strategies and measures, including strengthening equipment and improving drainage systems, to enhance the substation's flood prevention and disaster resistance capabilities;

[0114] The substation area is divided into three flood prevention levels:

[0115] Level 1 flood control area: For highly sensitive areas, including transformer areas and control rooms / dispatching rooms, flood control measures will be taken when the waterlogging depth reaches 10 cm below the bottom of the base. For these areas, flood control measures include: strengthening equipment waterproofing measures to improve the waterproof performance of equipment; strengthening drainage systems to ensure smooth drainage and timely removal of waterlogged water;

[0116] Secondary flood control areas: including high-voltage transmission areas and switchgear areas, flood control measures will be taken when the waterlogging depth reaches 20 cm below the bottom of the basement; for these areas, flood control measures include: strengthening equipment waterproofing measures to improve equipment's ability to resist floods; strengthening monitoring systems, real-time monitoring of waterlogging conditions, and taking timely response measures;

[0117] Level 3 flood control area: including auxiliary facilities area and maintenance / repair area. When the depth of waterlogging reaches 30 cm below the bottom of the base, flood control measures will be taken. For these areas, flood control measures include: strengthening the waterproof performance of equipment to ensure the normal operation of equipment; strengthening personnel on duty, promptly detecting waterlogging and taking measures to deal with it.

[0118] Beneficial effects of the present invention:

[0119] 1. In the existing technical means, waterlogging and accumulation are usually predicted in areas where waterlogging often occurs to achieve early warning, but to a certain extent, the uncontrollability of waterlogging occurrence areas at different times is ignored. The present invention uses machine learning technology to build a combined model (RF-Informer model) to predict the location of waterlogging in substations and the changes in waterlogging depth over time, provide comprehensive spatiotemporal prediction information, and achieve a comprehensive assessment of waterlogging risks.

[0120] 2. The present invention makes comprehensive use of large-scale multi-source data, including rainfall data, terrain data, hydrological data, etc., extracts key features, and establishes a complex prediction model, thereby more accurately predicting the possibility and extent of waterlogging.

[0121] 3. The model involved in the present invention has the ability to efficiently process large-scale data, realize real-time prediction and emergency response, and can quickly provide early warning information when waterlogging occurs, helping power grid operators to take preventive measures in a timely manner.

[0122] 4. The present invention takes into account the different requirements for flood control in different areas of the substation. The model can provide personalized flood control strategies and measures in a targeted manner, providing more targeted support for substation waterlogging prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0123] Figure 1 Build a diagram for the spatiotemporal prediction model of waterlogging;

[0124] Figure 2 A schematic diagram of the technical route of a method for spatiotemporal prediction of substation waterlogging based on machine learning;

[0125] Figure 3 This is the probability map of the location where waterlogging occurs. DETAILED DESCRIPTION

[0126] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0127] like Figure 1-2 As shown in FIG. 1 , a method for spatiotemporal prediction of substation waterlogging based on machine learning includes the following steps:

[0128] S1, data collection and preprocessing;

[0129] S2, construct substation waterlogging coupling model;

[0130] S3, design different rainfall schemes to generate rainfall-waterlogging samples;

[0131] S4, construct a spatial prediction model for waterlogging;

[0132] S5. Construct a time series prediction model for waterlogging;

[0133] S6, model combination and optimization;

[0134] S7, Model evaluation and application;

[0135] S8. Develop targeted flood prevention and treatment measures.

[0136] The S1 specifically includes:

[0137] 1.1、Data Collection:

[0138] First, remote sensing technology and LiDAR scanning are used to obtain surface elevation data of the substation area to provide detailed terrain information of the substation and its surrounding areas to ensure accurate simulation of water flow paths and waterlogging locations. Next, detailed information such as the layout and flow of the drainage network in the substation area is collected to understand the capacity and efficiency of the drainage system and provide basic data for simulating drainage and waterlogging. In addition, data such as the spatial distribution, structural characteristics, and usage of roads and buildings around the substation are collected to identify obstacles that may affect water flow and waterlogging, providing accurate spatial information for the hydrodynamic model. Finally, boundary condition data, including meteorological data (such as rainfall, rainfall intensity, wind speed, etc.) and hydrological data (such as river water level, flow, etc.), are collected to provide external driving conditions for the hydrodynamic model to ensure the authenticity and reliability of the simulation results.

[0139] 1.2 Data processing and integration:

[0140] In the data cleaning stage, statistical analysis and graphic visualization methods are used to detect and remove noise and erroneous values, and to fill in missing data. For surface elevation data, it is necessary to remove abnormal elevation points and use interpolation methods to fill in missing elevation data; for drainage network data, it is necessary to correct incorrect pipe diameter and material information to ensure the consistency of flow data. These steps are aimed at improving the accuracy and completeness of the data and ensuring the high quality of the model input data.

[0141] During the format conversion process, data from different sources are converted into a unified file format (CSV). The original LiDAR point cloud data is converted into a digital elevation model (DEM), and the drainage network map in PDF format is converted into GIS data in vector format. This process ensures consistency between different data sets and facilitates data integration and processing.

[0142] Finally, all data were converted to a unified coordinate system (WGS84) to ensure spatial matching between different datasets. The surface elevation data was converted from the local coordinate system to the WGS84 coordinate system, and the drainage network layout data was aligned to the unified coordinate system to ensure consistency with other spatial data. The unified coordinate system processing step ensured spatial matching between different datasets and avoided deviations caused by inconsistent coordinates.

[0143] Through the above data processing and integration steps, various types of collected data are strictly processed and integrated to ensure that they can be smoothly imported into the model for simulation, laying a solid foundation for building an accurate MIKE 21 hydrodynamic model.

[0144] The S2 specifically includes:

[0145] The purpose of using the MIKE 21 hydrodynamic model to construct a substation waterlogging coupling model is to accurately simulate and predict the hydrological and water resources conditions in the substation area.

[0146] First, the model was constructed using MIKE 21 software. MIKE 21 is a professional hydrodynamic simulation software suitable for complex water flow system analysis. The key to model construction is to prepare the necessary input data, including surface elevation data, drainage network information, spatial distribution data of roads and buildings, and boundary condition data. These data will directly affect the accuracy of the model and the simulation effect.

[0147] Secondly, the mathematical description of the model is based on the water flow continuity equation, the water flow motion equation and the bed friction slope, and the formula is as follows:

[0148] Continuity equation for water flow:

[0149]

[0150] Where h is the water depth, u is the depth-averaged velocity along the x-direction, v is the depth-averaged velocity along the y-direction, t is time, and x and y are spatial coordinates.

[0151] The equation of motion for water flow is:

[0152] The momentum equation in the x direction is:

[0153]

[0154] The momentum equation in the y direction is:

[0155]

[0156] Where g is the acceleration due to gravity, S bx and S by are the components of the bed slope in the x and y directions, S fx and S fy are the components of the bed friction slope in the x and y directions respectively.

[0157] Bed surface friction slope:

[0158] The bed surface friction slope is usually calculated using the Manning formula:

[0159]

[0160] Where n is the Manning roughness coefficient;

[0161] MIKE 21 solves these equations numerically, taking into account the influence of terrain, and can accurately simulate the propagation and evolution of water flow under complex terrain and structures.

[0162] Next, the parameter calibration process is carried out. Parameter calibration is to adjust the internal parameters of the model by comparing with the actual observation data so that the simulation results are as consistent as possible with the measured data. This process usually relies on historical rainfall events and determines the best parameter combination by minimizing the difference between the model prediction value and the measured value. The method usually used is to minimize the residual sum of squares:

[0163]

[0164] Among them, M i is the model prediction value, O i This process involves repeatedly adjusting the model parameters until the model's predictions and the measured data achieve satisfactory consistency and accuracy.

[0165] Furthermore, the model is validated. The validation process tests the predictive ability of the parameterized model by using independent rainfall events. The purpose of the validation is to confirm whether the model can stably predict the hydrodynamic response in the substation area under different rainfall conditions. The validation results can evaluate the reliability and applicability of the model.

[0166] Through the above steps, a two-dimensional hydrodynamic model based on MIKE 21 was constructed, and its accuracy and reliability in simulating the hydrodynamic response under different rainfall conditions were ensured through parameter calibration and verification.

[0167] The S3 specifically includes:

[0168] In the rainfall scheme design stage, the key step is to set rainfall schemes of different intensities and durations according to the required rainfall scenarios, drive the MIKE 21 hydrodynamic model to simulate, and generate rainfall-flooding samples under different rainfall conditions.

[0169] 3.1 Rainfall scheme design:

[0170] Select rainfall type and characteristics: Select the appropriate rainfall type, such as continuous light rain, heavy rain, or intermittent rainfall, based on the climate and hydrological characteristics of the region.

[0171] Set rainfall intensity and duration: Set the rainfall intensity R0 and duration τ. These parameters drive the rainfall process through mathematical models and affect the flow and accumulation of surface water.

[0172] The exponential decay model of rainfall intensity is expressed as:

[0173]

[0174] Where R(t) is the rainfall intensity at time t, R0 is the initial rainfall intensity, and τ is the time constant, describing the decay rate of rainfall intensity over time.

[0175] Spatial distribution characteristics: Considering the spatial distribution characteristics of rainfall in the study area, such as uniform rainfall, concentrated rainfall or local rainfall, different distribution patterns will have a significant impact on surface runoff and waterlogging risks.

[0176] 3.2 Simulate and generate rainfall-waterlogging samples:

[0177] The designed rainfall scheme was input into the MIKE 21 hydrodynamic model for numerical simulation to obtain rainfall-waterlogging sample data under different rainfall conditions, which served as the basis for subsequent model training and verification.

[0178] The S4 specifically includes:

[0179] In this step, the location of waterlogging in the substation area is predicted by constructing a spatial waterlogging prediction model. This process mainly includes data set construction, establishment of a spatial waterlogging prediction model, and result visualization.

[0180] 4.1 Dataset Construction:

[0181] First, the grid waterlogging spatial feature data and simulation result data are obtained to form a training data set. The grid waterlogging spatial feature data includes information such as terrain, drainage network, roads and buildings, while the simulation result data comes from the simulation results of the MIKE 21 model under different rainfall conditions. These data are combined to form a data set (part of the data is shown in Table 1), and preprocessed, including data cleaning, format conversion, feature extraction and normalization. The preprocessed data set is divided into a training set (80%) and a validation set (20%) for model training and validation.

[0182] Table 1 Part of the input data for training RF

[0183]

[0184] Normalization formula:

[0185]

[0186] The normalized data range is [0, 1], where min(x) and max(x) are the minimum and maximum values ​​of the sample data, respectively.

[0187] 4.2 Establishment of spatial prediction model for waterlogging:

[0188] A spatial prediction model for substation waterlogging is established based on the Random Forest (RF) algorithm. The Random Forest algorithm constructs multiple decision trees and makes predictions through a majority voting mechanism, which has high accuracy and robustness.

[0189] The random forest algorithm is an ensemble learning method that combines the results of multiple decision trees to make predictions, thereby improving the accuracy and stability of the model. During training, each decision tree randomly extracts a subset from the sample data and randomly selects some features in the feature space for node splitting. The final prediction result is the average or majority vote of all tree prediction results.

[0190] Decision tree construction:

[0191] For each decision tree T i , the training data set is randomly sampled to obtain a subset D i , randomly select a subset F from the feature set F i The decision tree construction process is as follows:

[0192] T i =BuildTree(D i ,F i )

[0193] Among them, BuildTree is the decision tree construction algorithm, D i is a sample subset, F i is a feature subset.

[0194] A collection of trees:

[0195] A random forest contains N decision trees, forming a collection of trees:

[0196] RF={T1,T2,…,T N}

[0197] predict:

[0198] For a new input feature vector X, each tree T i Give a prediction result P i , the prediction results of all trees are obtained by majority voting or averaging to get the final result P:

[0199]

[0200] Among them, P is the probability of waterlogging, which ranges from 0 to 1. The closer it is to 1, the greater the possibility of waterlogging. X is the input feature vector, including terrain, drainage facilities, buildings, roads and other features.

[0201] During the model training process, the preprocessed training data set is used to train the random forest model to predict the probability of waterlogging in different areas within the substation.

[0202] 4.3 Results Visualization:

[0203] In order to more intuitively show the possibility of waterlogging, the color mapping technology is used to visualize the prediction results. The specific method is to map the predicted waterlogging probability value to a color gradient (value between 0-1, such as Figure 3 Different colors represent different probabilities of waterlogging. The closer the value is to 1, the greater the probability of waterlogging in the area. This can intuitively show the waterlogging risks at different locations within the substation area.

[0204] The S5 specifically includes:

[0205] In this step, by constructing a waterlogging time series prediction model, the temporal variation of waterlogging depth in different areas of the substation is predicted. This process mainly includes data set construction and establishment of a waterlogging time series prediction model.

[0206] 5.1 Dataset Construction:

[0207] First, the prediction results of the waterlogging location are obtained from the waterlogging spatial prediction model, and combined with the simulation results of the corresponding grid waterlogging points, a waterlogging time series data set is generated (part of the data is shown in Table 2), and it is preprocessed, including data cleaning, format conversion, feature extraction and normalization. The preprocessed data set is divided into a training set (80%) and a validation set (20%) for model training and validation.

[0208] Table 2 Part of the input data of Informer

[0209] Grid point number Time Waterlogging depth (cm) 1 2024-05-01 12:00:00 20 1 2024-05-01 12:10:00 30 1 2024-05-01 12:20:00 40 … … … 2 2024-05-01 12:00:00 15 2 2024-05-01 12:10:00 18 2 2024-05-01 12:20:00 20 … … … 3 2024-05-01 12:00:00 30 3 2024-05-01 12:10:00 35 3 2024-05-01 12:20:00 50 4 2024-05-01 12:00:00 20 4 2024-05-01 12:10:00 25 4 2024-05-01 12:20:00 30 … … …

[0210] 5.2 Establishment of waterlogging time series prediction model:

[0211] In order to build a waterlogging time series prediction model, the Informer model is used. This model is based on the attention mechanism and has efficient computing power and good prediction performance. The design goal of the Informer model is to improve the computational efficiency and accuracy of long time series prediction through self-attention mechanism, multi-layer network stacking and step-by-step decoding method.

[0212] The Informer model is a network structure based on the attention mechanism, which mainly improves the computational efficiency of the self-attention mechanism, multi-layer network stacking and step-by-step decoding methods. The model is mainly composed of an encoder and a decoder. The encoder receives a large amount of long sequence input, and uses a multi-head sparse self-attention mechanism to replace the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and fills the target elements with 0 values.

[0213] The encoder consists of multiple multi-head sparse self-attention and distillation. The multi-head sparse self-attention mechanism is a variant of the self-attention mechanism. It is used to capture the dependencies between features. Its advantage is that it reduces the computational complexity of each layer of attention. The multi-head sparse self-attention mechanism achieves O(LlogL) in terms of time complexity and memory usage, and has comparable performance for sequence dependency alignment. Its calculation formula is:

[0214]

[0215] Where: Q, K and V are matrices obtained by different linear transformations of the input variables; is the matrix obtained after Q is probabilistically sparse; d is the dimension of the query vector and the key-value vector; Softmax is the activation function.

[0216] Multi-head sparse self-attention is composed of several stacked sparse self-attention layers. Generally speaking, there are multiple multi-head sparse self-attention layers in the encoder. The distillation operation is performed when the input data is transferred from the i-th multi-head sparse self-attention layer to the i+1-th multi-head sparse self-attention layer. The purpose of the distillation operation is to reduce the complexity of the model and reduce the amount of calculation, giving higher weights to the main features. The main process of extracting features from the j-th layer to the j+1-th layer will be simplified as follows to speed up the operation. After encoding the input sequence, the encoder generates a concatenated feature map and sends it to the multi-head self-attention layer of the decoder.

[0217]

[0218] Where: [*] is the attention module in multi-head sparse self-attention; Conv1d represents the regularized convolution operation on the time series; ELU is the activation function; MaxPool represents the maximum pooling operation.

[0219] The "distillation" operation adds a pooling layer with a stride of 2, making Downsample to 1 / 2, thereby reducing memory utilization to O((2–λ)LlogL), where the value of λ is very small. To enhance the robustness of the distillation operation, the Informer model constructs 1 copy by halving the input and reduces the number of attention blocks to align the output dimensions. The decoder part mainly includes the Masked multi-head sparse self-attention block and the multi-head attention block. Its attention mechanism is to prevent each position from paying attention to future information and avoid autoregression. The Masked multi-head sparse self-attention block receives the label vector of the following formula, and the decoder uses a generative method for decoding. The decoder uses the intermediate result output by the encoder to perform a multi-head attention operation, adjusts the dimension of the output data through the fully connected layer, and finally outputs the result of the predicted sequence length in one go.

[0220]

[0221] Where: is the t-th input sequence of the decoder; is the start mark of the tth sequence; is a placeholder for the target sequence of the tth sequence.

[0222] According to the location of waterlogging predicted by the random forest algorithm, combined with the corresponding grid simulation results, a waterlogging time series data set is generated. This data set includes the changes in waterlogging depth in different areas over time, that is, the data on waterlogging depth in different areas of the substation at different times. A waterlogging time prediction model is established based on the Informer time series model. The generated waterlogging time series data is used as input to predict the changing trend of waterlogging depth in different areas of the substation over time, so as to take preventive measures and respond to waterlogging risks in a timely manner.

[0223] The S6 specifically includes:

[0224] 6.1 Model Combination:

[0225] The spatial prediction model of waterlogging based on the Random Forest (RF) algorithm is combined with the temporal prediction model of waterlogging based on the Informer time series model. The spatiotemporal prediction model of waterlogging is constructed as follows: Figure 1 As shown in the figure, the sequential series method is used to optimize the combined model to improve the prediction accuracy and practicality.

[0226] First, the RF model is used to predict the spatial distribution of waterlogging. The RF model constructs multiple decision trees and makes predictions through a majority voting mechanism, which has high accuracy and robustness.

[0227] Secondly, the waterlogging locations predicted by the RF model are combined with the simulation results of the corresponding grid waterlogging points to generate a waterlogging time series dataset. This dataset is input into the Informer model to predict the change of waterlogging depth over time in different areas of the substation.

[0228] Through this sequential cascade method, we can first determine the high-risk areas for waterlogging through the RF model, and then use the Informer model to predict the changing trend of waterlogging depth in these areas over time, thereby achieving more accurate spatiotemporal waterlogging predictions.

[0229] 6.2 Optimization Model:

[0230] Parameter tuning is an important means to optimize model performance. By adjusting the model's hyperparameters, the model's prediction accuracy and robustness can be significantly improved.

[0231] Parameter tuning of random forest (RF) model:

[0232] The random forest model is an integrated model composed of multiple decision trees, and its performance is affected by multiple hyperparameters. Commonly used hyperparameters and their tuning methods are as follows:

[0233] Number of trees: The more trees there are, the more stable the model is, but the computational cost is also higher. You can use grid search or random search to find the optimal value within a certain range.

[0234] Maximum tree depth: Controls the maximum depth of each tree to prevent overfitting. Deeper trees can capture complex patterns but are prone to overfitting. The optimal depth can be selected through cross-validation.

[0235] Parameter tuning of the Informer model:

[0236] The Informer model is a complex deep learning model whose performance is affected by multiple hyperparameters. Commonly used hyperparameters and their tuning methods are as follows:

[0237] The number of encoder and decoder layers: The more layers there are, the stronger the model's expressiveness is, but it is prone to overfitting. The optimal number of layers can be selected through cross-validation.

[0238] Number of attention heads: The number of heads in the multi-head attention mechanism. Increasing the number of heads can capture more features, but the amount of computation also increases. The optimal number of heads can be selected through experiments.

[0239] The S7 specifically includes:

[0240] The validation set data is used to evaluate the model under the optimal parameters, and the prediction accuracy of the model is evaluated by comparing it with the actual observed data. The mean square error (MSE), mean absolute error (MAE), and goodness of fit R 2 As model evaluation indicators, these metrics can help us understand the model's predictive ability and error level.

[0241]

[0242] Where n is the number of prediction points; y i are the predicted value and true value of the i-th point data in the test data set respectively; is the average of the observed values.

[0243] After the evaluation is completed, the optimal combined model will be applied to the actual substation waterlogging prediction to provide accurate support and decision-making basis for substation waterlogging management and disaster prevention and mitigation.

[0244] The S8 specifically includes:

[0245] According to the characteristics of different areas of the substation and different requirements for waterlogging, corresponding flood prevention and treatment strategies and measures are formulated, including strengthening equipment, improving drainage systems, etc., to improve the substation's flood prevention and disaster resistance capabilities.

[0246] The substation area is divided into three flood prevention levels:

[0247] Level 1 flood control area: For highly sensitive areas, such as transformer areas and control rooms / dispatching rooms, flood control measures will be taken when the waterlogging depth reaches 10 cm below the bottom of the base. For these areas, flood control measures may include: strengthening equipment waterproofing measures to improve the waterproof performance of equipment; strengthening drainage systems (using portable submersible pumps) to ensure smooth drainage and timely removal of waterlogged water.

[0248] Secondary flood control areas: including high-voltage transmission areas and switchgear areas, flood control measures will be taken when the waterlogging depth reaches 20 cm below the bottom of the base. For these areas, flood control measures may include: strengthening equipment waterproofing measures to improve equipment's ability to resist floods; strengthening monitoring systems to monitor waterlogging in real time and taking timely response measures.

[0249] Level 3 flood control area: including auxiliary facilities area and maintenance / repair area, when the waterlogging depth reaches 30 cm below the bottom of the base, flood control measures will be taken. For these areas, flood control measures may include: strengthening the waterproof performance of equipment to ensure normal operation of equipment; strengthening personnel on duty to promptly detect waterlogging and take measures to deal with it.

[0250] By dividing the substation area into different flood control levels and taking corresponding flood control measures for each level, we can more effectively respond to waterlogging incidents and ensure the safe and stable operation of substation facilities.

[0251] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for spatiotemporal prediction of substation waterlogging based on machine learning, characterized in that: It includes the following steps: S1, data collection and preprocessing; S2, construct substation waterlogging coupling model; S3, design different rainfall schemes to generate rainfall-waterlogging samples; S4, construct a spatial prediction model for waterlogging; S5. Construct a time series prediction model for waterlogging; S6, model combination and optimization; S7, Model evaluation and application; S8. Develop targeted flood prevention and treatment measures.

2. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1 is characterized by: The S1 specifically includes: S1.

1. Data collection: First, the surface elevation data of the substation area is scanned to provide detailed terrain information of the substation and its surrounding areas to ensure accurate simulation of water flow paths and waterlogging locations. Next, detailed information on the layout and flow of the drainage network in the substation area is collected to understand the capacity and efficiency of the drainage system and provide basic data for simulating drainage and waterlogging. In addition, the spatial distribution, structural characteristics and usage data of roads and buildings around the substation are collected to identify obstacles that may affect water flow and waterlogging, providing accurate spatial information for the hydrodynamic model. Finally, boundary condition data, including meteorological data and hydrological data, are collected to provide external driving conditions for the hydrodynamic model and ensure the authenticity and reliability of the simulation results. S1.2 Data processing and integration: In the data cleaning stage, statistical analysis and graphic visualization methods are used to detect and remove noise and error values, and to fill in missing data. For surface elevation data, abnormal elevation points are removed, and interpolation methods are used to fill in missing elevation data. For drainage network data, incorrect pipe diameter and material information are corrected to ensure the consistency of flow data. During the format conversion process, data from different sources are converted into a unified file format; the original LiDAR point cloud data is converted into a digital elevation model, and the drainage network map in PDF format is converted into GIS data in vector format. This process ensures the consistency between different data sets and facilitates data integration and processing; Finally, all data are converted to the unified WGS84 coordinate system to ensure spatial matching between different data sets; the surface elevation data is converted from the local coordinate system to the wGs84 coordinate system, and the drainage network layout data is aligned to the unified coordinate system to ensure consistency with other spatial data.

3. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1 is characterized by: The S2 specifically includes: First, the model was constructed using MIKE 21 software; necessary input data were prepared, including surface elevation data, drainage network information, spatial distribution data of roads and buildings, and boundary condition data; Secondly, the mathematical description of the model is based on the water flow continuity equation, the water flow motion equation and the bed friction slope, and the formula is as follows: Continuity equation for water flow: Where h is the water depth, u is the depth-averaged velocity along the x direction, v is the depth-averaged velocity along the y direction, t is time, and x and y are spatial coordinates; The equation of motion for water flow is: The momentum equation in the x direction is: The momentum equation in the y direction is: Where g is the acceleration due to gravity, S bx and S by are the components of the bed slope in the x and y directions, S fx and S fy are the components of the bed surface friction slope in the x and y directions, respectively; Bed surface friction slope: The bed friction slope is usually calculated using the Manning formula: Where n is the Manning roughness coefficient; Parameter calibration process: Parameter calibration is to adjust the internal parameters of the model by comparing with the actual observation data, so that the simulation results are as consistent as possible with the measured data; this process usually relies on historical rainfall events and determines the best parameter combination by minimizing the difference between the model prediction value and the measured value; the method used is to minimize the residual sum of squares: Among them, M i is the model prediction value, O i is the corresponding measured value; this process involves repeatedly adjusting the model parameters until the model's prediction results reach satisfactory consistency and accuracy with the measured data; Conduct model validation: The validation process tests the predictive ability of the parameterized model by using independent rainfall events. The purpose of the validation is to confirm whether the model can stably predict the hydrodynamic response in the substation area under different rainfall conditions. The validation results can evaluate the reliability and applicability of the model. Through the above steps, a two-dimensional hydrodynamic model based on MIKE 21 was constructed, and its accuracy and reliability in simulating the hydrodynamic response under different rainfall conditions were ensured through parameter calibration and verification.

4. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1 is characterized by: The S3 specifically includes: S3.

1. Rainfall scheme design: Select rainfall type and characteristics: Select appropriate rainfall type according to the climate and hydrological characteristics of the region; Setting rainfall intensity and duration: Setting the rainfall intensity R0 and duration τ. These parameters drive the rainfall process through mathematical models and affect the flow and accumulation of surface water. The exponential decay model of rainfall intensity is expressed as: Among them, R(t) is the rainfall intensity at time t, R0 is the initial rainfall intensity, and τ is the time constant, which describes the decay rate of rainfall intensity over time; Spatial distribution characteristics: Considering the spatial distribution characteristics of rainfall in the study area, different distribution patterns will have a significant impact on surface runoff and waterlogging risks; S3.

2. Simulate and generate rainfall-waterlogging samples: The designed rainfall scheme was input into the MIKE 21 hydrodynamic model for numerical simulation to obtain rainfall-waterlogging sample data under different rainfall conditions, which served as the basis for subsequent model training and verification.

5. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1 is characterized by: The S4 specifically includes: S4.

1. Dataset construction: First, the grid waterlogging spatial feature data and simulation result data are obtained to form a training data set; the grid waterlogging spatial feature data includes terrain, drainage network, road and building information, while the simulation result data comes from the simulation results of the MIKE 21 model under different rainfall conditions. These data are combined to form a data set, which is preprocessed, including data cleaning, format conversion, feature extraction and normalization steps; the preprocessed data set is divided into a training set (80%) and a validation set (20%) for model training and validation; Normalization formula: The normalized data range is [0, 1], where min(x) and max(x) are the minimum and maximum values ​​of the sample data, respectively; S4.

2. Establishment of spatial prediction model for waterlogging: A spatial prediction model for substation waterlogging was established based on the random forest algorithm; Decision tree construction: For each decision tree T i , the training data set is randomly sampled to obtain a subset D i , randomly select a subset F from the feature set F i Conduct training; the process of building a decision tree is as follows: T i =BuildTree(D i ,F i ) Among them, BuildTree is the decision tree construction algorithm, D i is a sample subset, F i is a feature subset; A collection of trees: A random forest contains N decision trees, forming a collection of trees: RF={T1,T2,...,T N } predict: For a new input feature vector X, each tree T i Give a prediction result P i , the prediction results of all trees are obtained by majority voting or averaging to get the final result P: Among them, P is the probability of waterlogging, which ranges from 0 to 1. The closer it is to 1, the greater the possibility of waterlogging. X is the input feature vector, including terrain, drainage facilities, buildings, and road features. During the model training process, the preprocessed training data set was used to train the random forest model to predict the probability of waterlogging in different areas of the substation; S4.

3. Result visualization: In order to more intuitively display the possibility of waterlogging, color mapping technology is used to visualize the prediction results.

6. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1, characterized in that: The S5 specifically includes: S5.

1. Dataset construction: Firstly, the prediction results of waterlogging locations are obtained from the waterlogging spatial prediction model, and combined with the simulation results of the corresponding grid waterlogging points, a waterlogging time series dataset is generated and preprocessed, including data cleaning, format conversion, feature extraction and normalization steps; the preprocessed dataset is divided into a training set (80%) and a validation set (20%) for model training and validation; S5.

2. Establishment of waterlogging time series prediction model: In order to build a waterlogging time series prediction model, the Informer model is used. The design goal of the Informer model is to improve the computational efficiency and accuracy of long time series prediction through self-attention mechanism, multi-layer network stacking and step-by-step decoding method. The Informer model includes an encoder and a decoder. The encoder receives a large amount of long sequence input and uses a multi-head sparse self-attention mechanism to replace the conventional self-attention mechanism. The distillation operation is used to extract the main attention to reduce the network size. The decoder accepts long sequence input and fills the target elements with 0 values. The encoder consists of multiple multi-head sparse self-attention and distillation; the calculation formula of the multi-head sparse self-attention mechanism is: Where: Q, K and V are matrices obtained by different linear transformations of the input variables; is the matrix obtained after Q is probabilistically sparse; d is the dimension of the query vector and the key-value vector; Softmax is the activation function; The multi-head sparse self-attention is composed of several stacked sparse self-attention layers; there are multiple multi-head sparse self-attention layers in the encoder; when the input data is transferred from the i-th multi-head sparse self-attention layer to the i+1-th multi-head sparse self-attention layer, a distillation operation is performed; the purpose of the distillation operation is to reduce the complexity of the model and reduce the amount of calculation, giving higher weights to the main features, and the process of extracting features from the j-th layer to the j+1-th layer will be simplified as follows to speed up the operation; after encoding the input sequence, the encoder generates a concatenated feature map and sends it to the multi-head self-attention layer of the decoder; Where: [*] is the attention module in multi-head sparse self-attention; Conv1d represents the regularized convolution operation on the time series; ELU is the activation function; MaxPool represents the maximum pooling operation; The "distillation" operation adds a pooling layer with a stride of 2, making Downsample to 1 / 2, thereby reducing memory utilization to O((2–λ)LlogL), where the value of λ is very small; to enhance the robustness of the distillation operation, the Informer model constructs one copy by halving the input, and reduces the number of attention blocks to align the output dimensions; the decoder part includes Masked multi-head sparse self-attention blocks and multi-head attention blocks. The attention mechanism is to prevent each position from paying attention to future information and avoid autoregression; the Masked multi-head sparse self-attention block receives the label vector of the following formula, and the decoder uses a generative method for decoding; the decoder uses the intermediate result output by the encoder to perform a multi-head attention operation, adjusts the dimension of the output data through the fully connected layer, and finally outputs the result of the predicted sequence length at one time; Where: is the t-th input sequence of the decoder; is the start mark of the tth sequence; is the placeholder for the target sequence of the tth sequence; According to the location of waterlogging predicted by the random forest algorithm, a waterlogging time series data set is generated in combination with the corresponding grid simulation results. This data set includes the changes in the waterlogging depth in different areas over time, that is, the data on the waterlogging depth in different areas of the substation at different times; a waterlogging time prediction model is established based on the Informer time series model, and the generated waterlogging time series data is used as input to predict the changing trend of the waterlogging depth in different areas of the substation over time, so as to take preventive measures and respond to waterlogging risks in a timely manner.

7. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1, characterized in that: The S6 specifically includes: S6.

1. Model combination: The spatial waterlogging prediction model based on the Random Forest (RF) algorithm is combined with the temporal waterlogging prediction model based on the Informer time series model. The combined model is optimized using a sequential concatenation method to improve prediction accuracy and practicability. First, the RF model is used to predict the spatial distribution of waterlogging; the RF model constructs multiple decision trees and makes predictions through a majority voting mechanism; Secondly, the waterlogging locations predicted by the RF model are combined with the simulation results of the corresponding grid waterlogging points to generate a waterlogging time series dataset. This dataset is input into the Informer model to predict the change of waterlogging depth over time in different areas of the substation. Through this sequential cascade method, the high-risk areas of waterlogging are first determined by the RF model, and then the changing trend of waterlogging depth in these areas over time is predicted by the Informer model, thus achieving more accurate spatiotemporal waterlogging prediction; S6.

2. Optimization model: Parameter tuning of random forest (RF) model: The random forest model is an integrated model composed of multiple decision trees. Its performance is affected by multiple hyperparameters. The hyperparameters and their tuning methods are as follows: Number of trees: The more trees there are, the more stable the model is, but the computational cost is also higher; use grid search or random search to find the optimal value within a certain range; Maximum depth of the tree: Control the maximum depth of each tree to prevent overfitting; deeper trees can capture complex patterns but are prone to overfitting; select the optimal depth through cross-validation; Parameter tuning of the Informer model: The hyperparameters and their tuning methods are as follows: The number of layers of the encoder and decoder: The more layers there are, the stronger the model's expressiveness is, but it is prone to overfitting; the optimal number of layers is selected through cross-validation; Number of attention heads: The number of heads in the multi-head attention mechanism. Increasing the number of heads can capture more features, but the amount of computation also increases. The optimal number of heads is selected through experiments.

8. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1, characterized in that: The S7 specifically includes: The validation set data was used to evaluate the model under the optimal parameters, and the prediction accuracy of the model was evaluated by comparing it with the actual observed data. The mean square error (MSE), mean absolute error (MAE), and goodness of fit R 2 As a model evaluation indicator; Where n is the number of prediction points; y i are the predicted value and true value of the i-th point data in the test data set respectively; is the mean of the observed values; After the evaluation is completed, the optimal combined model will be applied to the actual substation waterlogging prediction to provide accurate support and decision-making basis for substation waterlogging management and disaster prevention and mitigation.

9. The method for spatiotemporal prediction of substation waterlogging based on machine learning according to claim 1, characterized in that: The S8 specifically includes: According to the characteristics of different areas of the substation and the different requirements for waterlogging, formulate corresponding flood prevention and treatment strategies and measures, including strengthening equipment and improving drainage systems, to enhance the substation's flood prevention and disaster resistance capabilities; The substation area is divided into three flood prevention levels: Level 1 flood control area: For highly sensitive areas, including transformer areas and control rooms / dispatching rooms, flood control measures will be taken when the waterlogging depth reaches 10 cm below the bottom of the base. For these areas, flood control measures include: strengthening equipment waterproofing measures to improve the waterproof performance of equipment; strengthening drainage systems to ensure smooth drainage and timely removal of waterlogged water; Secondary flood control areas: including high-voltage transmission areas and switchgear areas, flood control measures will be taken when the waterlogging depth reaches 20 cm below the bottom of the basement; for these areas, flood control measures include: strengthening equipment waterproofing measures to improve equipment's ability to resist floods; strengthening monitoring systems, real-time monitoring of waterlogging conditions, and taking timely response measures; Level 3 flood control area: including auxiliary facilities area and maintenance / repair area. When the depth of waterlogging reaches 30 cm below the bottom of the base, flood control measures will be taken. For these areas, flood control measures include: strengthening the waterproof performance of equipment to ensure the normal operation of equipment; strengthening personnel on duty, promptly detecting waterlogging and taking measures to deal with it.

Citation Information

Cited By

  • Peak regulation demand evaluation system for electric power system

    CN119401667A

  • Terrain-adaptive rainstorm waterlogging spatio-temporal joint prediction method and system

    CN121350714A

  • Terrain-adaptive storm waterlogging spatio-temporal joint prediction method and system

    CN121350714B