Water diversion channel ice condition intelligent prediction method based on artificial intelligence

By combining GRA, XGBoost, QPSO and LightGBM algorithms, the key factors of ice conditions are identified and model parameters are optimized, and the existing problems of low accuracy of ice conditions prediction and long debugging time are solved, achieving more efficient and accurate The ice situation prediction ensures the safe operation of the water diversion channel in winter.

CN119066538APending Publication Date: 2024-12-03TIANJIN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411200424.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prediction accuracy of the existing ice situation prediction model is not high, the prediction period is short, the model debugging time is long, and the applicability and accuracy under different conditions are difficult to guarantee.

Method used

Using an artificial intelligence-based method, combined with GRA, XGBoost, QPSO and LightGBM algorithms, we identify key influence factors of ice conditions and perform adaptive intelligent rate determination of model parameters to achieve intelligent prediction of ice conditions. Specific steps include identification and quantification of key ice conditions, input factor preprocessing, initialization of ice conditions prediction model and optimization decision-making, and output ice conditions classification prediction results.

Benefits of technology

It improves the accuracy and efficiency of ice situation prediction, shortens the model debugging time, reduces the error caused by the hyperparameter setting of the model, ensures the safety of the water diversion channel in winter, and has broad application prospects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066538A_ABST
    Figure CN119066538A_ABST
Patent Text Reader

Abstract

The invention discloses a water diversion channel ice condition intelligent prediction method based on artificial intelligence. Ice condition key influence factors are identified and quantified, input factors and ice condition classification are preprocessed, normalized input factors and initial ice condition states are input into a LightGBM model, and the input factors are classified into various ice conditions; an optimal input factor combination for ice condition prediction is obtained iteratively through a QPSO algorithm, input factors of the ice condition prediction model are updated, after iteration is finished, particles with the optimal fitness value serve as optimal solutions, and optimal hyper-parameters are obtained; applying the optimal hyper-parameter to a LightGBM model, searching for an optimal segmentation node for an input factor node, inputting an optimized input factor decision result into an ice condition prediction model, carrying out ice condition classification prediction, and outputting ice condition states including an ice-free state, a shore ice state, a drifting ice state and an ice cover state. According to the invention, high-precision and high-efficiency intelligent prediction of the ice condition of the diversion channel is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of water conservancy project informatization, and in particular to an intelligent prediction method for ice conditions in a water diversion channel. Background Art

[0002] The ice condition prediction model for water diversion channels is one of the important technical means to simulate and predict the changes in ice and icicles in water transfer and diversion projects, and is also one of the important non-engineering measures to reduce ice and icicle disasters in water conservancy projects under complex winter environments. Under cold winter climate conditions, ice and icicles may cause blockages in water diversion channels, affect the smooth flow of water, and even damage the structure of water diversion channels. The ice condition prediction model is widely used in the scientific prevention and control of ice and icicle disasters and ensuring the safety of water transfer during the ice period. It has certain practical significance for protecting people's lives and property and the safety of water transfer during the ice period. Accurate and efficient ice condition prediction models can predict the formation and development of ice and icicles in advance, help formulate scientific prevention and control measures, and ensure the safe and smooth operation of water transfer and diversion projects.

[0003] At present, when conducting ice forecasting research, analysis and research are often carried out through prototype observations, experimental research and numerical simulation, that is, simulation calculations are performed by assuming parameters. However, scholars currently mainly use methods such as mathematical statistics and numerical simulation. Due to the complexity of the modeling process and the limitations of observational data, there is a lack of more systematic and intelligent research, and the accuracy and efficiency of the model need to be further improved. In addition, the selection and adjustment of model parameters usually require a lot of trial and error processes. The complexity and amount of calculation make it difficult to obtain the global optimal prediction results. When applied on a large scale, the workload and time cost are huge, and the applicability and accuracy of the model under different conditions are difficult to guarantee.

[0004] With the continuous development of artificial intelligence technology, many advanced intelligent algorithms have been applied to solve the problem of ice condition change prediction and have achieved good results. GRA method, XGBoost method, QPSO algorithm and LightGBM model are important branches of artificial intelligence technology. GRA is a multi-factor statistical analysis method with a small demand for samples. It can eliminate subjectivity in system analysis to a certain extent and takes up little memory space during the calculation process. XGBoost is an integrated learning method based on decision trees, which has high efficiency and accuracy and is widely used in hydrology, machinery and other fields. The QPSO algorithm applies quantum theory to particle behavior, updates the particle position based on the average optimal particle position of the quantum particle group and uses the shrinkage and expansion factor to avoid the particle swarm algorithm from easily falling into the local optimal problem and enhances the global convergence ability. The main features are the group search strategy and the information exchange between groups, which can usually solve the problem in a time significantly lower than that required by conventional calculation methods. The LightGBM model adopts improved methods such as histogram algorithm and leaf growth strategy with depth restriction, which significantly improves the training speed of the model. The model has the characteristics of fast training speed, low memory usage, excellent performance in processing large-scale data, and support for parallel and distributed training. At the same time, it has high accuracy and good scalability, and is suitable for processing big data problems such as classification and prediction. Intelligent algorithms also have significant advantages in input factor determination, model parameter optimization methods and prediction accuracy. In terms of determining input factors, intelligent algorithms can automatically identify and extract key factors that affect changes in ice conditions, such as water temperature, air temperature, negative accumulated temperature, etc. The algorithm significantly reduces the interference of human subjective judgment through a data-driven approach; in terms of model parameter optimization, intelligent algorithms can gradually adjust model parameters through continuous iteration and self-learning to achieve the optimal state; in terms of model prediction accuracy and efficiency, the adaptive optimization process of the intelligent prediction model not only improves the prediction accuracy of the model, but also significantly improves efficiency, avoiding the tedious trial and error process in traditional methods. Summary of the invention

[0005] In view of the technical problems of low prediction accuracy, short forecast period and long model debugging time in existing ice condition prediction models, the present invention proposes an intelligent prediction method for ice conditions in water diversion channels based on artificial intelligence, aiming to identify and quantify key influencing factors of ice conditions through GRA method and XGBoost method, and realize adaptive intelligent calibration of ice condition prediction model parameters by combining QPSO algorithm with LightGBM model, so as to realize intelligent prediction of ice conditions in water diversion channels.

[0006] In order to achieve the above-mentioned object of the invention, the present invention proposes the following technical solutions:

[0007] The present invention proposes an artificial intelligence-based intelligent prediction method for ice conditions in a water diversion channel, comprising the following steps:

[0008] Identify and quantify the key influencing factors of ice conditions: Use the Grey Relational Analysis (GRA) algorithm to identify the correlation sequence with large correlation values ​​between water temperature and ice conditions and meteorological environmental factors as the key influencing factors; Use the correlation sequence with high correlation as the feature variable, and select the quantitative feature variables with high importance as the input factor variables of the prediction model based on the decision tree ensemble learning XGBoost algorithm;

[0009] Preprocessing the input factors and ice condition classification: normalizing the input factors, and inputting the normalized input factors and initial ice condition status into the LightGBM model;

[0010] Initialize the ice prediction model: use the QPSO algorithm to iteratively obtain the optimal input factor combination for ice prediction, and update different combinations of specific configuration parameters of multiple historical key meteorological environmental factors of the input factors of the ice prediction model. After the iteration, the particle with the best fitness value is the optimal solution, and the input factor parameter corresponding to the particle is the optimal hyperparameter;

[0011] The optimal hyperparameters are applied to the LightGBM model, and the optimal split nodes are found for the input factor nodes. The optimal split nodes are combined to realize the input factor optimization decision. The optimized input factor decision results are input into the ice condition prediction model to perform ice condition classification prediction. The output predicted ice conditions include ice-free state, shore ice state, drift ice state and ice cover state.

[0012] Compared with the prior art, the present invention can achieve the following beneficial technical effects:

[0013] 1. The prediction method, which combines the advantages of GRA, XGBoost, QPSO and LightGBM (GRA-XGBoost-QPSO-LightGBM), can accurately identify the factors affecting ice conditions, optimize the prediction model parameters through a simplified calibration process of model parameters, reduce the error caused by model hyperparameter settings, shorten the model debugging time, and improve the accuracy and efficiency of ice condition prediction, so as to better ensure the safety of water diversion channel operation in winter;

[0014] 2. It has important practical significance and broad application prospects in practical applications, and can provide effective technical support for ice age prediction of water diversion channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the overall process of an intelligent prediction method for ice conditions in a water diversion channel based on artificial intelligence of the present invention;

[0016] Figure 2This is a schematic diagram of the internal structure of LightGBM;

[0017] Figure 3 It is a schematic diagram of the histogram algorithm;

[0018] Figure 4 This is an example diagram of the grey correlation coefficient results between influencing factors and ice conditions. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0020] The present invention proposes an intelligent prediction method for ice conditions in a water diversion channel based on artificial intelligence.

[0021] like Figure 1 As shown, the present invention is an intelligent prediction method for ice conditions in a water diversion channel based on artificial intelligence, and the specific process is as follows:

[0022] Step 1: Identify and quantify the key factors affecting ice conditions; for example, the key factors affecting winter ice conditions include water temperature, air temperature (minimum temperature, daily average temperature), three-day sliding negative accumulated temperature, three-day sliding air temperature, air pressure, relative humidity, wind speed, ground temperature, sunshine time, flow, etc. The key factors affecting ice clearance are used as input factors of the prediction model; wherein:

[0023] The grey correlation analysis GRA algorithm is used to conduct grey correlation analysis on the key influencing factors of ice conditions. The specific steps include dimensionless processing of the data of the key influencing factors of ice conditions. The data of each influencing factor is dimensionless processed to the interval [0,1] to eliminate the influence of different dimensions between the influencing factors; the actual observation data of ice conditions are selected as the original sequence to determine the key influencing factors of ice conditions, and the correlation between influencing factors such as meteorology and flow and ice conditions is analyzed; the ice conditions, water temperature, minimum temperature, average temperature, maximum temperature, negative accumulated temperature, sliding three-day negative accumulated temperature, The grey correlation coefficient between the sequence of air pressure, relative humidity, wind speed, ground temperature, sunshine time and flow and the ice condition sequence; the grey correlation coefficient between the quantitative influencing factors and the ice state and the grey influence degree on the evolution of ice conditions are identified, and the evaluation results are the characteristic importance values ​​of each influencing factor for four ice condition classifications, namely, no ice is set to 0, shore ice is set to 1, drift ice is set to 2, and ice cover is set to 3; the correlation sequence established based on the original sequence is the sequence of water temperature, minimum temperature, average temperature, maximum temperature, negative accumulated temperature, sliding three-day negative accumulated temperature, air pressure, relative humidity, wind speed, ground temperature, sunshine time and flow.

[0024] Specifically, a grey relational analysis (GRA) of key factors affecting ice conditions was performed to obtain the correlation between GRA quantitative water temperature and ice conditions and various meteorological environmental factors such as air temperature, extreme cold waves, negative accumulated temperature, daily average air pressure, daily average wind speed, daily average humidity, ground temperature, flow, and water level. The process is as follows:

[0025] The original sequence composed of the actual observation data of GRA quantified ice conditions is determined to be Q0(i) = [Q0(1), Q0(2), ..., Q0(n)], i = 1, 2, 3, ..., n; where n is the length of the sequence, and i represents the number of various key factors (meteorological environmental factors); the associated sequence of the original sequence composed of GRA quantified water temperature, minimum temperature, average temperature, maximum temperature, negative accumulated temperature, sliding three-day negative accumulated temperature, air pressure, relative humidity, wind speed, ground temperature, sunshine time and flow is determined to be Q k (i) = [Q k (1),Q k (2),...,Q k (n)], k = 1, 2, 3, ..., m, where m is the length of the association sequence;

[0026] The original sequence and the associated sequence are dimensionless, and the expression of the initialized original sequence Q'0(i) is as follows:

[0027]

[0028] Then we get the normalized initialization original sequence Q'0(i) and the normalized initialization correlation sequence Q' k (i);

[0029] Calculate the initialization original sequence Q'0(i) and the initialization associated sequence Q' k The absolute difference Δ of (i) k (i), the expression is as follows:

[0030] Δ k (i)=|Q'0(i)-Q' k (i)| (2)

[0031] Determine the absolute difference Δ k The maximum absolute value of the difference between the initialization original sequence of the i-th key factor and the initialization associated sequence of the i-th key factor in (i) max and the minimum absolute value of the difference Δ min , the expression is as follows:

[0032] Δ max =max k max i |Q'0(i)-Q' k (i)| (3)

[0033] Δ min =min k min i |Q'0(i)-Q' k (i)| (4)

[0034] Then calculate the correlation coefficient of each initialized correlation sequence and the initialization original sequence corresponding to the i-th key factor, the expression is as follows:

[0035]

[0036] Where ρ is the resolution coefficient, its value is in the interval (0,1), usually ρ=0.5;

[0037] Finally, calculate each initialization association sequence Q' k (i) The correlation degree r of the initialized original sequence Q'0(i) k , the expression is as follows:

[0038]

[0039] According to the correlation r k Determine the correlation between each associated sequence and the original sequence. The larger the correlation value, the greater the influence of the corresponding associated sequence on the original sequence, and vice versa.

[0040] Grey correlation analysis has a small demand for samples and can eliminate subjectivity in system analysis to a certain extent. At the same time, it does not take up much memory space during the calculation process, so it is widely used in many research fields. This method uses GRA to quantify the degree of correlation between water temperature and ice conditions and various meteorological environmental factors such as air temperature, extreme cold waves, negative accumulated temperature, daily average air pressure, daily average wind speed, daily average humidity, ground temperature, flow, water level, etc., and compares it with the XGBoost calculation results to further determine the key influencing factors of water temperature and ice conditions; next, the ensemble learning algorithm based on the decision tree is used to quantify the importance of the influencing factor characteristics. The specific process is described as follows;

[0041] The decision tree uses the extreme gradient boosting tree XGBoost. The XGBoost model can handle both regression problems and feature selection problems. Set the relevant parameters of the extreme gradient boosting tree XGBoost: define the leaf nodes of the tree model in the XGBoost model as the classification labels of ice condition prediction. The XGBoost objective function consists of a loss function and a regularization term. The loss function is used to measure the gap between the ice condition classification prediction and the actual situation, while the regularization term is used to prevent the model from overfitting. By optimizing the objective function, the key meteorological environmental factors that affect water temperature and ice conditions can be effectively identified and quantified. XGBoost quantifies the importance of feature variables through feature selection. The higher the score of the feature variable, the more important the feature is. The highly correlated association sequences are used as features, and they are arranged in descending order according to the feature importance scores to obtain the feature variables with the highest importance, so as to further screen the key influencing factors. Thus, through XGBoost, the factors that are important to the water temperature and ice conditions are evaluated from various meteorological environmental factors such as air temperature, extreme cold waves, negative accumulated temperature, daily average air pressure, daily average wind speed, daily average humidity, ground temperature, flow, water level, etc., as key influencing factors, and the weights of the influencing factors are determined to provide a basis for the prediction of water temperature and ice conditions. By gradually improving the performance of each tree in the XGBoost model, optimizing the split point of the tree model, quantifying the contribution of key influencing factors to ice condition prediction, the changes in ice conditions can be predicted more accurately.

[0042] Output results of decision tree-based ensemble learning algorithm It is expressed by formula (7):

[0043]

[0044] Where K represents the total number of decision trees, k∈K, F={f(x)=ω q(x)}, f k (x j ) represents the classification result of the jth sample in the kth tree, f(x) represents the prediction function of a single decision tree, the mapping from feature space to prediction results, ω q(x) represents the weight of the leaf node, which represents the predicted value of sample xxx on the corresponding leaf node. ω represents the weight parameter of the leaf node, that is, the predicted value on each leaf node. x represents the feature vector of the sample, which contains the values ​​of all key factors. q(x) represents the structural function of the decision tree, which maps the input feature x to the corresponding leaf node index. represents a set of real numbers, representing the space where the input features or output prediction values ​​are located. Represents the T-dimensional real vector space, that is, the space where each sample feature vector x is located.

[0045] Calculate the decision tree gain as shown below:

[0046]

[0047] In the formula, I=I L ∪I R , I L Represents the sample of the left node space of the decision tree, I R represents the samples in the node space on the right side of the decision tree, I represents all samples in the node space of the decision tree, λ and γ represent regularization parameters, and g j 、h j They represent the first-order gradient and the second-order gradient respectively, and i represents the sample number;

[0048] XGBoost determines the best split node of the tree model of key factors affecting ice conditions through decision tree gain. In order to further optimize the selection of split nodes and more accurately predict ice condition changes, the best split point initially determined by XGBoost will be optimized again in the subsequent steps in combination with the histogram algorithm. The histogram algorithm is used to obtain the best split point of the final tree model of key factors affecting ice conditions. The feature importance is determined based on the average gain, which significantly improves the computational efficiency and accuracy of ice condition classification prediction. Specifically, the more times a feature is split, the greater the corresponding decision tree model gain, and the more important the feature is;

[0049] Step 2: perform data preprocessing, including normalizing the input factors so that they are between [0, 1] to improve the convergence efficiency of the machine learning algorithm (LightGBM model) based on the gradient boosting framework; input the input factors and the corresponding four types of ice conditions (no ice is set to 0, shore ice is set to 1, drift ice is set to 2, and ice cover is set to 3) into the LightGBM model; initialize the machine learning algorithm (LightGBM model) based on the gradient boosting framework with min_child_samples=2, max_depth=2, learning_rate (learning rate)=0.1, and num_leaves=2. Among them, min_child_samples indicates the minimum number of samples that each leaf node must contain when classifying ice conditions, which helps ensure that the model does not overfit to a few abnormal situations when dealing with small sample sizes, thereby enhancing the stability of the prediction results; max_depth indicates the maximum depth of the decision tree, which limits the depth of the tree to prevent the model from over-learning the noise in the data when facing complex ice conditions, thereby improving the model's generalization ability for different ice conditions; num_leaves indicates the maximum number of leaf nodes in the decision tree, which controls the complexity of the tree together with max_depth, and ensures that the model can concisely and effectively identify and classify different types of ice conditions when dividing the input meteorological and hydrological factors, thereby improving the prediction accuracy of the model;

[0050] Step 3: Initialize the ice prediction model, that is, initialize the parameters of the quantum-behavior particle swarm optimization algorithm (QPSO algorithm), including the particle population size, the maximum number of iterations, and the coefficient β that controls the particle convergence speed. The input of the ice prediction model includes various meteorological environmental factors such as water temperature, air temperature, extreme cold waves, negative accumulated temperature, daily average air pressure, daily average wind speed, daily average humidity, ground temperature, flow, water level, etc. The output is the ice state, including four categories: no ice (0), shore ice (1), drift ice (2), and ice cover (3). Specifically, the particle population represents the configuration of the ice prediction model represented by each particle in the QPSO algorithm, that is, a set of model parameters related to the ice evolution process; the population size represents the number of particles considered simultaneously in the ice prediction optimization process. A larger population size means that the model can explore different ice evolution paths in a wider parameter space; the particle position represents the specific configuration of multiple key meteorological environmental factors in the model, such as different combinations of parameters such as temperature, wind speed, and flow. By iteratively optimizing these parameter combinations, the model can more accurately predict future ice conditions; the coefficient β that controls the particle convergence speed is used to adjust the agility of the ice prediction model in finding the optimal parameter combination, ensuring that the model can not only stably predict common ice conditions, but also flexibly respond to complex ice changes caused by extreme weather.

[0051] The QPSO algorithm applies quantum theory to particle behavior and improves the particle swarm algorithm in two main aspects. On the one hand, the particle's movement speed is no longer considered, and the particle position is updated according to the average optimal particle position of the quantum particle swarm, which avoids the particle swarm algorithm from falling into the local optimal problem and enables particles to search for the global optimal solution more efficiently. On the other hand, the contraction and expansion factor is used to control the update of the particle position, which enhances the global convergence ability. The specific principles are as follows:

[0052]

[0053] P i (t+1)=α·pbest i (t)+(1-α) gbest j (t) (10)

[0054]

[0055] In the formula, x i represents the particle position of the i-th particle, N represents the total number of particles, and pbest i represents the average position of the i-th particle, mbest represents the historical average best position of the particle, and gbest jrepresents the position of the best particle j in the entire population, α represents a random number uniformly distributed in [0,1], and β represents the contraction and expansion factor, which can control the rapid convergence of particles and avoid the local swing of particles at the optimal value, so that the algorithm can more easily find the global optimal solution. Among them, t represents the current number of iterations. Each iteration model updates the input parameters such as temperature, wind speed, water level, etc., and gradually adjusts the parameter configuration through multiple iterations to find the most accurate factor combination for ice condition prediction; u represents the influencing factor of the global optimal solution, which is related to the optimal ice condition prediction model parameters found in the current population, and is used to guide other particles to move to a better ice condition prediction configuration to improve the overall prediction accuracy; t i represents the number of updates of the historical best position of the ith particle. In each iteration, each particle updates its historical best position and optimizes the parameter configuration of the model in predicting specific ice conditions (such as shore ice and drift ice); P i (t) represents the current position of the i-th particle in the t-th iteration, that is, the specific configuration of the current model parameters. By continuously updating the position of the particles, the ice condition prediction model is gradually optimized to more accurately reflect the changes in ice conditions under actual meteorological and hydrological conditions;

[0056] Step 4: Perform fitness evaluation of the QPSO algorithm model, that is, calculate the mean square error (MSE) of the fitness function of the current QPSO particle as the fitness function:

[0057]

[0058] In the formula, represents the ice-clear classification prediction value of the mth iteration, y m Represents the true value of the ice-clear classification at the mth iteration.

[0059] The fitness of the QPSO algorithm model is evaluated according to the fitness function MSE; if the evaluation result is that the optimal condition, i.e., the fitness threshold, is not met, step five is executed;

[0060] Step 5: Update the particle position, and gradually adjust the parameter configuration through multiple iterations of the model to find the most accurate parameter combination for ice forecasting. The particle position represents the different combinations of specific configuration parameters of multiple key meteorological environmental factors in the model. The updated particle position moves to a better ice forecasting configuration. Through multiple iterations of the model, the parameter configuration is gradually adjusted to find the most accurate parameter combination for ice forecasting. The model can more accurately predict future ice conditions.

[0061] Step 6: Determine the hyperparameters of the ice condition prediction model, repeat steps 3 to 4, and update the particle position until the maximum number of iterations is reached. After the iteration, use the fitness function to calculate the optimal position and global optimal position of each particle in the particle swarm. The value of the global particle's optimal position gbest is the optimal value sought. The optimal solution is the particle with the best fitness value, and the parameters corresponding to the particle are the hyperparameters of the optimized LightGBM model.

[0062] Step 7: Apply the optimal hyperparameters to the LightGBM model and use the ice condition prediction model to perform ice condition prediction calculations.

[0063] The LightGBM is an upgrade algorithm based on the gradient boosting decision tree algorithm (GBDT) framework, and its internal structure is as follows Figure 2 As shown in the figure. The LightGBM model uses improved methods such as histogram algorithm and leaf-based growth strategy with depth limit, which significantly improves the training speed of the ice condition prediction model compared with the gradient boosting machine (GBM) model and the extreme gradient boosting (XGBoost) model. In big data application scenarios, when facing large sample and high-dimensional data sets, the LightGBM model has the characteristics of fast training speed, high computational efficiency and relatively accurate prediction, so it is particularly suitable for big data problems such as classification and prediction.

[0064] In this step, the histogram algorithm is used to discretize the continuous floating-point features into m units, and then these units are used to construct a histogram with a width of m; when the data is traversed, the statistical data is accumulated in the histogram with discrete values ​​as indexes; after the data is traversed once, the histogram accumulates the required statistics, and then the best segmentation point is found according to the discrete values ​​in the histogram, such as Figure 3 The optimal segmentation point initially determined needs to be optimized again in the subsequent steps in combination with the histogram algorithm. The final optimal segmentation point obtained by the histogram algorithm can further optimize the selection of segmentation nodes, significantly improving the computational efficiency and accuracy.

[0065] In addition to the histogram algorithm, LightGBM also proposes gradient-based single-side sampling (GOSS), exclusive feature binding (EFB), and leaf growing methods to further improve computational efficiency without compromising accuracy. When looking for the best segmentation in GOSS, the data is first sorted using the absolute value of the sample gradient. Then, data with higher gradients are retained, and data with lower gradients are randomly sampled. Finally, GOSS calculates the variance gain V of feature j j (d) to give the best split node.

[0066] Variance gain V j (d) is shown in the following formula:

[0067]

[0068] In the formula, A represents a sample data set with high gradient, B represents a sample data set with low gradient, and g i represents the gradient of each sample, a represents the sampling rate of high gradient data, b represents the sampling rate of low gradient data, and d represents the number of iterations.

[0069] The GOSS algorithm can accelerate model training by reducing samples, while the EFB algorithm uses the idea of ​​graph composition to treat features as nodes, connect non-mutually exclusive features, and then find all bundled feature sets. It can further reduce the data size by reducing the number of features.

[0070] The LightGBM base classifier uses a decision tree and is trained sequentially by fitting the negative gradient of the loss function. Its loss function can be obtained through a weighted combination scheme, as shown in formula (14):

[0071]

[0072] In the formula, M represents the maximum number of iterations, h m (x) represents the basic decision tree.

[0073] According to the optimal particle position, the model hyperparameters are set, and an intelligent ice condition prediction method based on GRA-XGBoost-QPSO-LightGBM is constructed to carry out intelligent ice condition prediction, conduct multi-classification prediction of ice conditions with a forecast period of 1d to 7d, and output the prediction results;

[0074] Step 8: Evaluation of prediction results. The classification accuracy of the prediction results of the method was analyzed and evaluated using indicators such as accuracy, recall, precision, F-score, Kappa coefficient, and ROC curve. At the same time, a comparative analysis was conducted with the QPSO-SVM and QPSO-RF methods to evaluate the accuracy of the ice condition intelligent prediction method based on GRA-XGBoost-QPSO-LightGBM.

[0075] According to the ice condition intelligent prediction method of the present invention, the QPSO algorithm is used to automatically optimize the min_child_samples, max_depth, learning_rate and num_leaves of the LightGBM model, where min_child_samples is the minimum number of samples in the leaf node; max_depth is the depth of the tree; learning_rate is the learning rate; num_leaves is the maximum leaf node tree. max_depth, learning_rate and num_leaves affect the classification accuracy of the model, min_child_samples can improve the generalization ability of the model, and the hyperparameters are optimized in the recommended value space by QPSO to improve the classification accuracy of the LightGBM model. The input data includes meteorological operating conditions, namely the minimum temperature, average daily temperature, flow, sliding three-day negative accumulated temperature, sliding three-day temperature and water temperature, and finally outputs the ice condition prediction result (no ice is set to 0, shore ice is set to 1, drift ice is set to 2, and ice cover is set to 3). According to the ice condition intelligent prediction method of the present invention, considering the influence of various influencing factors on ice condition changes, the ice condition intelligent prediction is carried out to forecast the ice condition prediction results.

[0076] The following is an example of the intelligent prediction of ice conditions at a typical measuring station to explain the present invention in detail and clearly. The present invention proposes an intelligent prediction method for ice conditions based on GRA-XGBoost-QPSO-LightGBM, the method comprising:

[0077] S1, identification and quantification of key factors affecting ice conditions;

[0078] Since water temperature, air temperature, negative accumulated temperature, sliding three-day negative accumulated temperature, sliding three-day air temperature, air pressure, relative humidity, wind speed, ground temperature, sunshine time, flow and other influencing factors all have an impact on the evolution of ice conditions, in order to identify and clarify key factors, the present invention introduces GRA and XGBoost to respectively identify the correlation between quantitative influencing factors and ice conditions and the degree of influence on the evolution of ice conditions, and evaluate the characteristic importance of each influencing factor to ice condition classification. The grey correlation coefficient between the ice conditions and influencing factors at the typical station of Beijuma is calculated by GRA, and the results are as follows: Figure 4 At the same time, XGBoost is used to quantify the feature importance of each influencing factor on ice condition classification. Table 1 shows the results of the feature importance analysis of ice condition influencing factors at Beijuma Station.

[0079] Table 1

[0080]

[0081] S2, data preprocessing and normalization;

[0082] From S1, we can know that the key factors affecting winter ice conditions are water temperature, minimum temperature, daily average temperature, sliding three-day negative accumulated temperature, sliding three-day temperature sum and flow. The input data are water temperature, minimum temperature, daily average temperature, sliding three-day negative accumulated temperature, sliding three-day temperature sum and flow of the typical station in Beijuma in the winter of 2020-2022, and they are normalized to be between [0,1] to improve the convergence efficiency of the LightGBM model. The key factors and the corresponding four types of ice conditions (no ice is set to 0, shore ice is set to 1, drift ice is set to 2, and ice cover is set to 3) are input into the LightGBM model. Table 2 shows the distribution of various types of ice conditions in winter at typical stations.

[0083] Table 2

[0084]

[0085] S3, model initialization;

[0086] Initialize the parameters of the QPSO algorithm, including the particle population size, the maximum number of iterations, and the coefficient that controls the particle convergence speed;

[0087] S4, model fitness evaluation;

[0088] Initialize the LightGBM model with min_child_samples = 2, max_depth = 2, learning_rate = 0.1, num_leaves = 2. Then, use the fitness function MSE to evaluate the fitness of the current particle;

[0089] S5, update particle position;

[0090] Calculate the average of the optimal positions of the particle swarm and update the particle positions.

[0091] S6, LightGBM model hyperparameters determination;

[0092] Repeat calculation steps three to four, update the particle position, and after the iteration, select the particle with the best fitness value as the optimal solution. The parameters corresponding to this particle are the hyperparameters of the optimized LightGBM model.

[0093] Step 7: Model calculation;

[0094] The optimal hyperparameters are applied to the LightGBM model. According to the optimal particle position, the model hyperparameters are set, and an intelligent ice condition prediction method based on GRA-XGBoost-QPSO-LightGBM is constructed. Intelligent ice condition prediction is carried out, and multi-classification prediction of ice conditions with a forecast period of 1d to 7d is performed, and the prediction results are output.

[0095] Step 8: Evaluation of prediction results.

[0096] The classification accuracy of the model prediction results was analyzed and evaluated using indicators such as accuracy, recall, precision, F-score, Kappa coefficient and ROC curve. At the same time, a comparative analysis was conducted with the QPSO-SVM and QPSO-RF methods to evaluate the accuracy of the ice condition intelligent prediction method based on GRA-XGBoost-QPSO-LightGBM. The receiver operating characteristic curve results of the ice condition intelligent prediction at the typical Beijuma station show that for the Beijuma station, the micro-average, macro-average and ROC curve area under the line AUC values ​​of the ice condition intelligent prediction method based on GRA-XGBoost-QPSO-LightGBM for each type of ice condition are 0.92, 0.92, 0.95, 0.88, 0.88 and 0.97, respectively, which are better than the corresponding 0.90, 0.88, 0.94, 0.81, 0.80 and 0.95 of QPSO-SVM and 0.90, 0.88, 0.94, 0.82, 0.83 and 0.95 of QPSO-RF. From the evaluation index, the ROC curve AUC value of the ice condition intelligent prediction method based on GRA-XGBoost-QPSO-LightGBM is the highest, showing a more excellent ice condition intelligent prediction performance.

[0097] As shown in Table 3, the results of the intelligent prediction of ice conditions in winter at the typical observation station of Beijuma using the intelligent prediction method of ice conditions in the present invention are shown. The results show that the intelligent prediction method of ice conditions based on GRA-XGBoost-QPSO-LightGBM shows excellent prediction performance in intelligent prediction of ice conditions.

[0098] Table 3

[0099]

[0100]

[0101] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications made without departing from the principle of the present invention should also be regarded as falling within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based intelligent prediction method for ice conditions in water diversion channels, characterized in that: The following steps are involved: Identify and quantify key factors affecting ice conditions: Use the Grey Relational Analysis (GRA) algorithm to identify correlation sequences with large correlation values ​​between water temperature and ice conditions and meteorological environmental factors as key influencing factors. Use the highly correlated correlation sequences as feature variables, and use the decision tree-based ensemble learning (XGBoost) algorithm to select highly important quantitative feature variables as input factor variables for the prediction model. Performing pre-processing on input factors and ice condition classification: normalizing the input factors; Initialize the ice prediction model: Use the QPSO algorithm to iteratively obtain the optimal input factor combination for ice prediction. Update different combinations of specific configuration parameters of multiple historical key meteorological environmental factors of the input factors of the ice prediction model. After the iteration, the particle with the best fitness value is the optimal solution, and the input factor parameters corresponding to this particle are the optimal hyperparameters. The optimal hyperparameters are applied to the LightGBM model, and the optimal split nodes are found for the input factor nodes. The optimal split nodes are combined to realize the input factor optimization decision. The optimized input factor decision results are input into the ice condition prediction model to perform ice condition classification prediction. The output predicted ice conditions include ice-free state, shore ice state, drift ice state and ice cover state.

2. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: The calculation process of the correlation coefficient between the water temperature and ice conditions and meteorological environmental factors is as follows: Calculate the initialization original sequence Q'0(i) and the initialization associated sequence Q' k The absolute difference Δ of (i) k (i), the expression is as follows: Δ k (i)=|Q'0(i)-Q' k (i)| Determine the absolute difference Δ k The maximum absolute value of the difference between the initialization original sequence of the i-th key factor and the initialization associated sequence of the i-th key factor in (i) max and the minimum absolute value of the difference Δ min , the expression is as follows: Δ max =max k max i |Q'0(i)-Q' k (i)| Δ min =my k my i |Q'0(i)-Q' k (in)| Then calculate the correlation coefficient of each initialized correlation sequence and the initialization original sequence corresponding to the i-th key factor, the expression is as follows: Where ρ is the resolution coefficient, Calculate each initialization association sequence Q' k (i) The correlation degree r of the initialized original sequence Q'0(i) k , the expression is as follows: Where n is the length of the sequence.

3. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: The fitness function is shown as follows: Where, represents the ice-clear classification prediction value of the mth iteration, y m Represents the true value of the ice-clear classification at the mth iteration.

4. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: The loss function of the LightGBM model is as follows: Where M represents the maximum number of iterations, h m (x) represents the basic decision tree.

5. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: The forecast results are multi-classification forecast values ​​of ice conditions for different forecast periods ranging from one to seven days.

6. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: Find the best split point based on the discrete values ​​in the histogram.

7. The method for intelligent prediction of ice conditions in water diversion channels based on artificial intelligence according to claim 1, characterized in that: Calculate the variance gain V of feature j according to GOSS j (d) Find the best split node.