Icing prediction feature selection method and device and computer equipment

Abstract

The invention relates to a feature selection method and device for icing prediction and computer equipment, and relates to the technical field of icing prediction. The method comprises the following steps: acquiring a plurality of historical icing samples; wherein the historical icing sample comprises a true value of an icing index and sample values of a plurality of icing characteristics, and the icing index comprises an icing thickness and / or an icing growth rate; for each icing feature, determining an importance index of the icing feature according to the true value in the same historical icing sample and the feature value of the icing feature; wherein the importance index at least comprises a characteristic use frequency and an average splitting gain; and according to the importance index of each icing feature, performing feature search on the plurality of icing features by using an improved sequence floating forward search method to obtain a key feature set for icing prediction. By adopting the method, a feature selection mechanism can be provided, and key features having a decisive effect on icing formation can be identified from massive icing features.

Description

Technical Field

[0001] The present application relates to the technical field of icing prediction, and particularly to a feature selection method, device and computer equipment for icing prediction. Background Art

[0002] In the power transmission network system, line icing is an important threat factor affecting the safe and stable operation of the power grid. Icing will significantly increase the weight of the wire, resulting in wire sag, swing and mechanical vibration. In severe cases, it can cause disasters such as wire breakage and tower collapse, resulting in large-scale power outages and huge economic losses. In recent years, the power department has invested a large amount of resources in line icing monitoring, established a comprehensive monitoring system including meteorological observation stations, online icing monitoring equipment, etc., and accumulated a large amount of multi-dimensional data, including meteorological factors such as temperature, humidity, wind speed, wind direction, precipitation, and relevant parameters such as line load, tower characteristics, and terrain features, providing a data basis for icing prediction and analysis.

[0003] However, existing icing prediction methods generally lack an effective feature selection mechanism and cannot identify key features that play a decisive role in icing formation from a large number of icing features. Summary of the Invention

[0004] Based on this, in view of the above technical problems, it is necessary to provide a feature selection method, device and computer equipment for icing prediction, which can provide a feature selection mechanism to identify key features that play a decisive role in icing formation from a large number of icing features.

[0005] In a first aspect, the present application provides a feature selection method for icing prediction, and the method includes:

[0006] Obtain a plurality of historical icing samples; wherein, the historical icing samples include the true values of icing indicators and the sample values of a plurality of icing features, and the icing indicators include icing thickness and / or icing growth rate;

[0007] For each icing feature, determine the importance index of the icing feature according to the true value and the feature value of the icing feature in the same historical icing sample; wherein, the importance index at least includes feature usage frequency and average splitting gain;

[0008] According to the importance index of each icing feature, use an improved sequential floating forward search method to perform feature search on a plurality of icing features to obtain a key feature set for icing prediction.

[0009] In one embodiment, for each icing feature, according to the true value in the same historical icing sample and the feature value of the icing feature, an importance index of the icing feature is determined, including: constructing an initial model for predicting the icing index and defining the objective function of each decision tree of the initial model; wherein, the objective function includes the true value of the icing index; for each decision tree of the initial model, according to the objective function of the decision tree, determine the split gain of each sample value of each icing feature as a feature split point, and according to the split gain, determine the split feature information of each node of the decision tree; according to the split feature information of each node of each decision tree, determine the importance index of each icing feature.

[0010] In one embodiment, for each decision tree of the initial model, according to the objective function of the decision tree, determine the split gain of each sample value of each icing feature as a feature split point, and according to the split gain, determine the split point information of each node of the decision tree, including: for each decision tree of the initial model, traverse each sample value of each icing feature in the historical icing sample located at the current node of the decision tree, and when traversing each sample value of each icing feature, according to the objective function of the decision tree, determine the split gain of the sample value of the icing feature; for each decision tree of the initial model, determine the target icing feature corresponding to the obtained maximum split gain and the target sample value of the target icing feature; for each decision tree of the initial model, according to the target icing feature and the target sample value, divide the historical icing sample located at the current node of the decision tree into the left child node and the right child node of the current node of the decision tree; for each decision tree of the initial model, in the case where the preset split condition is not satisfied, take the left child node and the right child node as the current nodes respectively, and return to execute the step of traversing each sample value of each icing feature in the historical icing sample located at the current node of the decision tree; for each decision tree of the initial model, in the case where the preset split condition is satisfied, obtain the constructed decision tree and determine the split point information of each node of the decision tree.

[0011] In one embodiment, before traversing each sample value of each icing feature in the historical icing sample located at the current node of each decision tree of the initial model, the method further includes: for each decision tree of the initial model, sort each sample value of each icing feature in the historical icing sample located at the current node of the decision tree in a specified order to obtain a sample value sequence of each icing feature; for each decision tree of the initial model, traversing each sample value of each icing feature in the historical icing sample located at the current node of the decision tree includes: for each decision tree of the initial model, traversing the sample value sequence of each icing feature.

[0012] In one embodiment, for each decision tree of the initial model, before taking the left child node and the right child node as the current node when the preset stop condition is not met, the method further includes: determining the loss value of dividing the samples with missing features into the left child node and the loss value of dividing the samples with missing features into the right child node; wherein, the samples with missing features are the samples in the historical icing samples of the current node of the decision tree that do not have the target icing feature; dividing the samples with missing features into the child node with the minimum loss value between the left child node and the right child node.

[0013] In one embodiment, the icing index includes the icing thickness and the icing growth rate; an initial model for predicting the icing index is constructed, and the objective function of each decision tree of the initial model is defined, including: constructing a first initial model for predicting the icing thickness and a second initial model for predicting the icing growth rate; defining the first objective function of each first decision tree of the first initial model and the second objective function of each second decision tree of the second initial model; for each decision tree of the initial model, according to the objective function of the decision tree, determining the split gain of each sample value of each icing feature as the feature split point, and according to the split gain, determining the split point information of each node of the decision tree, including: for each first decision tree, according to the first objective function of the first decision tree, determining the first split gain of each sample value of each icing feature as the feature split point, and according to the first split gain, determining the split feature information of each node of the first decision tree; for each second decision tree, according to the second objective function of the second decision tree, determining the second split gain of each sample value of each icing feature as the feature split point, and according to the first split gain, determining the split feature information of each node of the second decision tree; according to the split feature information of each node of each decision tree, determining the importance index of each icing feature, including: according to the split feature information of each node of each first decision tree, determining the first importance index of each icing feature, and according to the split feature information of each node of each second decision tree, determining the second importance index of each icing feature; for each icing feature, using a preset weight to perform weighted summation on the first importance index and the second importance index of the icing feature to obtain the importance index of the icing feature.

[0014] In one embodiment, according to the importance index of each icing feature, using an improved sequential floating forward search method, feature search is performed on multiple icing features to obtain a key feature set for icing prediction, including: sorting multiple icing features in descending order of feature usage frequency to obtain a first feature subset, and sorting multiple icing features in ascending order of average split gain to obtain a second feature subset; constructing a target feature set initially empty, a weighted evaluation function, and a comprehensive evaluation function; according to the sorting of each icing feature in the first feature subset and the second feature subset, in the first feature subset, search for the first icing feature that makes the function value of the weighted evaluation function of the target feature set increase and the increase value is the largest, and add the first icing to the target feature set; delete the first icing feature in the first feature subset to obtain a new first feature subset, and return to execute the step of searching in the first feature subset for the first icing feature that makes the function value of the weighted evaluation function of the target feature set the largest; in the case that there is no first icing feature in the first feature subset, according to the sorting of each icing feature in the target feature set in the first feature subset and the second feature subset, in the target feature set, search for the second icing feature that makes the function value of the comprehensive evaluation function of the target feature set increase after deletion; delete the second icing feature in the target feature set, and return to execute the step of searching in the target feature set for the second icing feature that makes the function value of the comprehensive evaluation function of the target feature set increase after deletion; determine that there is no second icing feature in the target feature set, and return to execute the step of searching in the first feature subset for the first icing feature that makes the function value of the weighted evaluation function of the target feature set the largest; in the case that the target feature set meets the preset stop condition, determine the current target feature set as the key feature set.

[0015] In one embodiment, before performing feature search on multiple icing features according to the importance index of each icing feature and using an improved sequential floating forward search method to obtain a key feature set for icing prediction, the method further includes: filtering zero-importance features for multiple icing features according to the feature usage frequency of each icing feature, and / or filtering redundancy for multiple icing features according to the mutual information between every two icing features to obtain an initial feature set; performing feature search on multiple icing features according to the importance index of each icing feature and using an improved sequential floating forward search method to obtain a key feature set for icing prediction, including: performing feature search on multiple icing features included in the initial feature set according to the importance index of each icing feature and using an improved sequential floating forward search method to obtain a key feature set for icing prediction.

[0016] In a second aspect, the present application also provides a feature selection device for icing prediction, including:

[0017] A sample acquisition module that acquires multiple historical icing samples; among them, the historical icing samples include the true values of icing indicators and the sample values of multiple icing characteristics, and the icing indicators include icing thickness and / or icing growth rate;

[0018] An index determination module, which is used to determine the importance index of each icing characteristic according to the true value in the same historical icing sample and the characteristic value of this icing characteristic; among them, the importance index at least includes the feature usage frequency and the average split gain;

[0019] A feature search module, which is used to perform feature search on multiple icing characteristics according to the importance index of each icing characteristic by using an improved sequential floating forward search method to obtain a key feature set for icing prediction.

[0020] In a third aspect, the present application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method embodiments in the first aspect are implemented.

[0021] In a fourth aspect, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method embodiments in the first aspect are implemented.

[0022] In a fifth aspect, the present application also provides a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the method embodiments in the first aspect are implemented.

[0023] The above-mentioned feature selection method, device and computer device for icing prediction obtain multiple historical icing samples, and each historical icing sample includes the true value of the icing indicator and the sample values of multiple icing characteristics. Thus, for each icing characteristic, the importance index of this icing characteristic can be determined according to the true value of the icing indicator in the same historical icing sample and the characteristic value of this icing characteristic. Furthermore, according to the importance index of each icing characteristic, an improved sequential floating forward search method is used to perform feature search on multiple icing characteristics to obtain a key feature set for icing prediction. In this way, on the one hand, a feature selection mechanism can be provided to use the improved sequential floating forward search method to identify the key features that play a decisive role in the formation of icing from multiple icing characteristics included in the historical icing samples; on the other hand, using the identified key features for icing prediction can improve the accuracy of the obtained prediction results and effectively improve the ability of the power grid to cope with icing disasters. Description of the Drawings

[0024] To more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the accompanying drawings required for the description of the embodiments of the present application or related technologies. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other related accompanying drawings can also be obtained based on these drawings.

[0025] Figure 1 It is a schematic flowchart of a feature selection method for icing prediction provided in some embodiments of the present application;

[0026] Figure 2 It is a schematic flowchart of determining a key feature set provided in some embodiments of the present application;

[0027] Figure 3 It is a schematic flowchart of a feature selection method for icing prediction provided in some other embodiments of the present application;

[0028] Figure 4 It is a schematic flowchart of determining the importance index of icing features provided in some embodiments of the present application;

[0029] Figure 5 It is a schematic flowchart of determining the importance index of icing features provided in some other embodiments of the present application;

[0030] Figure 6 It is a schematic flowchart of determining the splitting feature information of each node of each decision tree provided in some embodiments of the present application;

[0031] Figure 7 It is a schematic flowchart of a feature selection method for icing prediction provided in some other embodiments of the present application;

[0032] Figure 8 It is a structural block diagram of a feature selection device for icing prediction provided in some embodiments of the present application;

[0033] Figure 9 It is an internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0034] In order to make the objectives, technical solutions and advantages of the present application more clear, the following further details the present application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0035] In a power transmission network system, line icing is an important threat factor affecting the safe and stable operation of the power grid. Icing significantly increases the weight of the conductor, causing the conductor to sag, swing, and mechanically vibrate. In severe cases, it can lead to disasters such as wire breakage and tower collapse, resulting in large-scale power outages and huge economic losses. However, existing icing prediction methods generally lack an effective feature selection mechanism and are unable to identify key features that play a decisive role in ice formation from a vast amount of icing features.

[0036] Based on this, to solve the above technical problems, in an exemplary embodiment, a feature selection method for icing prediction is provided. This method can be applied to a computer device, which can be a server or a terminal. Among them, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Hereinafter, an example of applying this method to a server will be described. As Figure 1 shown, this method may include the following:

[0037] S101, obtain a plurality of historical icing samples.

[0038] Among them, the historical icing samples include the true values of icing indicators and the sample values of multiple icing features. The icing indicators include icing thickness and / or icing growth rate.

[0039] The so-called icing indicators are the data to be predicted during icing prediction. For example, icing thickness, icing growth rate, etc. The so-called icing features are multi-dimensional features related to and affecting the icing indicators. For example, features such as temperature, humidity, wind speed, precipitation in the meteorological dimension, and elevation, terrain features, vegetation coverage rate, etc. in the line parameter dimension. With the continuous development of icing prediction technology, a large amount of data in multiple dimensions can be obtained. For example:

[0040] 1. Time-series meteorological data: Temperature, humidity, wind speed, wind direction, precipitation, air pressure changes, etc. at multiple consecutive time points. [[ID=1 nine]]

[0041] 2. Line parameters: Conductor type, diameter, height, inclination angle, tension, etc. 3. Geographic information: Elevation, terrain features, vegetation coverage rate, surface conditions, etc. 4. Historical icing records: Historical icing thickness values and corresponding timestamps; growth values of icing thickness per unit time (icing growth rate). 5. Power operation data: Line load current, voltage, conductor surface temperature, etc. 6. Icing phase change parameters: Conversion rate from wet snow to ice, adhesion coefficient, etc.

[0042] Furthermore, through the analysis of the above-mentioned large amount of data, the icing prediction indicators required for icing prediction can be screened out, such as icing thickness and icing growth rate, as well as the icing characteristics for predicting the icing indicators, such as the data in 1, 2, 3, 5, and 6 above. Since the formation of icing is a complex physical-meteorological coupling process, involving numerous factors and non-linear interaction relationships, and the formation and development of icing have strong spatio-temporal differences, and the key influencing factors vary in different regions and different time periods, a feature selection mechanism can be provided to identify the key features that play a decisive role in the formation of icing from the large number of features in the above multi-dimensions.

[0043] Based on this, multiple historical icing samples can be obtained first. Each historical icing sample includes the true value of the icing indicator and the sample values of multiple icing characteristics. According to the different data required for icing prediction, the icing indicators of the historical icing samples can include icing thickness and / or icing growth rate.

[0044] Optionally, taking the icing indicators including icing thickness and icing growth rate as an example, multiple historical icing samples can be represented in the form of a time series table. Each row of the time series table represents a historical icing sample (in essence, an observation sample at a time point). Each column in the first N - 2 columns of the time series table represents an icing characteristic, and the last two columns represent icing thickness and icing growth rate respectively. Where N is the number of columns in the time series table, and N - 2 is the number of icing characteristics included in the historical icing sample.

[0045] Optionally, in some cases, in order to capture the time dependence of the icing formation process, among the multiple icing samples included in the historical icing samples, in addition to the icing characteristics that can be directly measured and collected by sensors and other devices, there can also be time series derivative characteristics constructed based on the directly obtained icing characteristics. For example, temperature change rate, temperature 3 hours ago, etc.

[0046] Exemplarily, for the j-th icing characteristic a j , its corresponding lagged characteristic a j,lag(τ) (t) = a j (t - τ) is constructed, where t represents the acquisition time of the icing characteristic a j , and τ represents the lag step. Furthermore, the change rate of the icing characteristics of key meteorological parameters (such as temperature, humidity, wind speed, etc.) is calculated as the time series derivative characteristic.

[0047] Specifically: Among them, Δa j (t) represents the time series derivative characteristic corresponding to the icing characteristic a j , δ is a preset time difference, and a j (t) represents the icing characteristic a collected at time tj The eigenvalue, a j (t - δ) represents the icing feature a collected at time t - δ j The eigenvalue of.

[0048] S102. For each icing feature, according to the true value in the same historical icing sample and the eigenvalue of this icing feature, determine the importance index of this icing feature.

[0049] Among them, the importance index includes at least the feature usage frequency and the average split gain.

[0050] In order to be able to identify the key features that play a decisive role in icing formation from a large number of icing features, as the key features for icing prediction, the importance index of each icing feature in the above historical icing sample can be determined, and the icing features can be selected according to the importance index.

[0051] For the key features that play a decisive role in icing formation, the correlation between this key feature and the icing index is relatively high. Therefore, in the same historical icing sample, the sample value of this key feature contributes more to the true value of the icing index. Therefore, for each icing feature, the importance index of this icing feature can be determined according to the true value in the same historical icing sample and the eigenvalue of this icing feature. Among them, the feature usage frequency (FScore) is used to represent the total number of times the icing feature is used as a split feature when constructing the decision tree of the icing prediction model, and the average split gain (AverageGain) is used to represent the average split gain brought by the icing feature as a split feature.

[0052] Optionally, the above importance index can also include the average cover (AverageCover) used to represent the number of historical icing samples covered on average when the icing feature is used as a split feature.

[0053] S103. According to the importance index of each icing feature, use the improved sequential floating forward search method to perform feature search on multiple icing features to obtain the key feature set for icing prediction.

[0054] After obtaining the importance index of each icing feature, according to the importance index of each icing feature, use the improved sequential floating forward search method to perform feature search on multiple icing features, so as to screen the key features that play a decisive role in icing formation from multiple icing features as the key features for icing prediction, thereby obtaining the key feature set.

[0055] Among them, the Improved Sequential Floating Forward Search (ISFFS) is a feature selection algorithm optimized based on the Sequential Floating Forward Search (SFFS). The so-called sequential floating forward search is a heuristic algorithm for feature selection, aiming to select the most representative and discriminative feature subset from a large number of candidate features to improve the performance of machine learning models, while reducing the complexity and computational cost of the models. It mainly includes three processes: initialization, feature addition phase, and feature floating phase. Specifically:

[0056] Initialization: Let the initial feature subset Put all n features into the candidate feature set C.

[0057] Feature addition phase: Select the i-th feature xi from the candidate feature set C such that after adding it to the current feature subset S, the evaluation index (such as classification accuracy, information gain, etc.) of the current feature subset reaches the optimal. Add xi to S and remove xi from C. Repeat this step until a certain stopping condition is met, such as reaching the preset number of features, or the evaluation index no longer improves significantly. Among them, i ∈ [1, n].

[0058] Feature floating phase: For each feature xj in the current feature subset S, try to remove it from S to obtain the temporary feature subset S' = S - {xj}. Evaluate the performance of S'. If the evaluation index of the temporary feature subset S' does not decrease or the decrease is within an acceptable range after removing a certain feature xj, then remove the feature xj from S formally. This step is to remove the redundant or less important features that may be introduced during the addition process.

[0059] Repeat the feature addition and feature floating phases until the algorithm converges, that is, no more features can be added or removed. At this time, the obtained feature subset S is the finally selected feature subset.

[0060] Furthermore, the improved sequential floating forward search not only starts from an empty set and gradually adds features like SFFS, but also starts from the full set and gradually deletes features under certain conditions. It explores the feature space more comprehensively through two-way search to avoid falling into local optima. For example, during the search process, it simultaneously maintains a feature subset that grows from an empty set and a feature subset that shrinks from the full set, and alternately adds and deletes features in both directions according to the evaluation index until the optimal feature subset is found. Thus, the improved sequential floating forward search can have better global search ability, more accurate feature selection, and stronger adaptability. In this way, the ice-covering features in the finally obtained key feature set can have a relatively high correlation with the ice-covering index, thereby improving the accuracy of the obtained key index. Furthermore, it can improve the accuracy of the prediction results obtained by using the identified key features for ice-covering prediction.

[0061] The above feature selection method for ice-covering prediction obtains multiple historical ice-covering samples, where each historical ice-covering sample includes the true value of the ice-covering index and the sample values of multiple ice-covering features. Thus, for each ice-covering feature, the importance index of this ice-covering feature can be determined according to the true value of the ice-covering index and the feature value of this ice-covering feature in the same historical ice-covering sample. Furthermore, according to the importance index of each ice-covering feature, the improved sequential floating forward search method is used to search for features among multiple ice-covering features to obtain the key feature set for ice-covering prediction. In this way, on the one hand, a feature selection mechanism can be provided to identify the key features that play a decisive role in ice-covering formation from the multiple ice-covering features included in the historical ice-covering samples by using the improved sequential floating forward search method; on the other hand, using the identified key features for ice-covering prediction can improve the accuracy of the obtained prediction results and effectively improve the power grid's ability to cope with ice-covering disasters.

[0062] Based on the above embodiments, in an exemplary embodiment, the determination of the key feature set in S103 is further refined. Optionally, as Figure 2 shown, it may include the following steps:

[0063] S201, sort the multiple ice-covering features in descending order of feature usage frequency to obtain the first feature subset, and sort the multiple ice-covering features in ascending order of average split gain to obtain the second feature subset.

[0064] As mentioned above, the importance index of the ice-covering feature includes feature usage frequency and average split gain. Thus, the multiple ice-covering features can be sorted in descending order of feature usage frequency to obtain the first feature subset, and the multiple ice-covering features can be sorted in ascending order of average split gain to obtain the second feature subset.

[0065] For example, the first feature subset is as follows:

[0066] I1 = {a1, a2,..., a m} where FScore combined (a1) ≥... ≥ FScore combined (a m )

[0067] where I1 represents the first feature subset, and FScore combined (a j ) represents the feature usage frequency of the j-th icing feature a j , where j ∈ [1, m] and m is the total number of icing features.

[0068] The second feature subset is as follows:

[0069] I2 = {a1, a2,..., a m} where AverageGain combined (a1) ≤... ≤ AverageGain combined (a m )

[0070] where I2 represents the second feature subset, and AverageGain combined (a j ) represents the average split gain of the j-th icing feature a j .

[0071] S202. Construct a target feature set, a weighted evaluation function, and a comprehensive evaluation function that are initially empty sets.

[0072] Construct a target feature set and initialize the target feature set as an empty set. Furthermore, construct a weighted evaluation function and a comprehensive evaluation function for the target feature set.

[0073] Optionally, when the icing index includes icing thickness and icing growth rate, the weighted evaluation function is as follows:

[0074] L(O) = γ · J th (O) + (1 - γ) · J rate (O)

[0075] where J(O) represents the weighted evaluation function, and L th (O) is the icing thickness prediction accuracy index using the target feature set; J rate (O) is the icing growth rate prediction accuracy index using the target feature set; γ is to balance J th (O) and J rate(O)'s weight coefficient is determined by the icing stage.

[0076] Optionally, in different formation stages of icing, the key features that are decisive for icing formation can be different. Thus, a dynamic weight for the icing stage can be introduced, that is, for the above weighted evaluation function, different γ can be set in different formation stages of icing. For example, as shown below:

[0077]

[0078] Among them, in the rapid ice growth stage, more attention is paid to the prediction of the ice growth rate. Optionally, in the rapid ice growth stage, γ rapia <0.5.

[0079] Optionally, the icing stage with an icing thickness below 5 mm and an ice growth rate within 0.1 mm / hour - 0.3 mm / hour can be regarded as the initial icing formation stage; the icing stage with an icing thickness within 5 mm - 15 mm and an ice growth rate within 0.3 mm / hour - 0.5 mm / hour can be regarded as the stable ice growth stage; the stage with an icing thickness exceeding 15 mm and an ice growth rate exceeding 0.5 mm / hour can be regarded as the rapid ice growth stage.

[0080] Optionally, the comprehensive evaluation function is as shown below:

[0081] J comprehensive (O) = J(O) + μJ crir (O)

[0082] Among them, J comprehensive (O) represents the comprehensive evaluation function, μ is the weight coefficient for identifying the importance of critical conditions, and J crit (O) represents the measurement function for measuring the effectiveness of the target feature set in identifying the critical conditions of icing formation and warning the rapid growth transition point.

[0083] S203. According to the sorting of each icing feature in the first feature subset and the second feature subset, in the first feature subset, search for the first icing feature that makes the function value of the weighted evaluation function of the target feature set increase and the increase value is the largest after adding, and add the first icing to the target feature set.

[0084] For the first feature subset, it is necessary to select icing features from the first feature subset and add them to the target feature set. Moreover, each time an icing feature is added to the target feature set, the function value of the weighted evaluation function of the target feature set is maximally improved. Specifically:

[0085] For each icing feature in the first feature set that may be added to the target feature set, calculate the function value of the weighted evaluation function of the target feature set after adding the icing feature according to the sorting of the icing feature in the first feature subset and the second feature subset, and calculate the change value of the weighted evaluation function before and after adding. Thus, select the icing feature that increases the function value of the weighted evaluation function of the target feature set and has the largest increase value as the first icing feature. Then, this first icing feature is the icing feature that makes the function value of the weighted evaluation function of the target feature set the largest after addition. Thus, the first icing feature can be added to the target feature set.

[0086] Optionally, in the case of introducing the dynamic weight in the icing formation stage, icing formation stage labels can be added to different icing features. Thus, for each icing feature in the first feature set that may be added to the target feature set, the weight value in the weighted evaluation function can be adjusted according to the formation stage label of the icing feature. Furthermore, according to the weighted evaluation function with the adjusted weight value and the sorting of the icing feature in the first feature subset and the second feature subset, calculate the function value of the weighted evaluation function of the target feature set after adding the icing feature, and calculate the change value of the weighted evaluation function before and after adding.

[0087] S204, Delete the first icing feature in the first feature subset to obtain a new first feature subset, and return to execute S203.

[0088] After adding the above first icing feature to the target feature set, the first icing feature in the first feature subset can be deleted. At this time, the first feature subset after deleting the first icing feature can be used as a new first feature subset. Thus, execute the above S203 again. In this way, through multiple loops, each icing feature in the first feature subset that can increase the function value of the weighted evaluation function of the target feature set can be added to the target feature set to continuously increase the function value of the weighted evaluation function of the target feature set.

[0089] S205, When there is no first icing feature in the first feature subset, search for a second icing feature in the target feature set that increases the function value of the comprehensive evaluation function of the target feature set after deletion according to the sorting of each icing feature in the target feature set in the first feature subset and the second feature subset.

[0090] Furthermore, after multiple cycles of the above S203 - S204, when it is impossible to search in the first feature subset for the first icing feature that, after being added, increases the function value of the weighted evaluation function of the target feature set and has the largest increase value, it can be stated that at this time, adding any icing feature in the first feature subset to the target feature set will reduce the function value of the weighted evaluation function of the target feature set. Then, there is no need to add the icing features in the first feature subset to the target feature set anymore. At this time, the forward addition stage in the improved sequential floating forward search stops and enters the backward screening stage. Specifically:

[0091] For each icing feature that may be deleted in the target feature set, according to the rankings of this icing feature in the first feature subset and the second feature subset, calculate the function value of the comprehensive evaluation function of the target feature set after deleting this icing feature, and calculate the change value of the comprehensive evaluation function before and after deletion. Thus, select the icing feature that increases the function value of the comprehensive evaluation function of the target feature set as the second icing feature. Then, this second icing feature is the icing feature that maximizes the function value of the comprehensive evaluation function of the target feature set after deletion. Thus, the second icing feature in the target feature set can be deleted.

[0092] S206, Delete the second icing feature in the target feature set and return to execute S205.

[0093] After determining the above second icing feature, the second icing feature in the target feature set can be deleted. At this time, the target feature set after deleting the second icing feature can be used as the new target feature set. Thus, execute the above S205 again. In this way, through multiple cycles, each icing feature in the target feature set that reduces the function value of the comprehensive evaluation function of the target feature set can be deleted, so that the function value of the comprehensive evaluation function of the target feature set can be continuously improved.

[0094] S207, Determine that there is no second icing feature in the target feature set and return to S203.

[0095] Furthermore, after multiple cycles of the above S205 - S206, when it is impossible to search in the target feature set for the second icing feature that increases the function value of the comprehensive evaluation function of the target feature set after deletion, it can be stated that at this time, deleting any icing feature in the target feature set will reduce the function value of the weighted evaluation function of the target feature set. Then, there is no need to delete the icing features in the target feature set anymore. At this time, the backward screening stage in the improved sequential floating forward search ends, and the forward addition stage can be entered again, that is, return to S203.

[0096] S208, When the target feature set meets the preset stop condition, determine the current target feature set as the key feature set.

[0097] By alternately performing the forward addition phase and the backward screening phase in the improved sequential floating forward search multiple times, the target feature set can gradually reach a preset stop condition. Thus, when the target feature set meets the preset stop condition, the current target feature set can be determined as the key feature set.

[0098] Optionally, the above preset stop condition can be that the target feature set does not change after two iterations, or it can be reaching a preset maximum number of iterations, or it can also be that the number of icing features included in the target feature set reaches a preset upper limit. Herein, one iteration means: performing the forward addition phase and the backward screening phase in the improved sequential floating forward search once.

[0099] In this embodiment, by using the ISFFS algorithm, a key feature set optimized for predicting the icing thickness and the icing growth rate can be obtained. Thus, on the one hand, among the multiple icing features included in the historical icing samples, the key features that play a decisive role in the formation of icing can be identified; on the other hand, using the identified key features for icing prediction can improve the accuracy of the obtained prediction results and effectively enhance the power grid's ability to cope with icing disasters.

[0100] In some cases, among the multiple icing features included in the historical icing samples, there may be some features with extremely low importance and redundancy. Therefore, in order to reduce the computational burden of subsequent feature search for multiple icing features using the improved sequential floating forward search method and improve the efficiency of icing feature selection, the multiple icing features included in the historical icing samples can be first subjected to feature filtering to delete features with extremely low importance and / or redundancy.

[0101] Based on this, on the basis of the above embodiments, in an exemplary embodiment, as Figure 3 shown, the feature selection method for icing prediction may include the following steps:

[0102] S301, obtain multiple historical icing samples.

[0103] S302, for each icing feature, determine the importance index of the icing feature according to the true value and the feature value of the icing feature in the same historical icing sample.

[0104] Among them, the specific implementation manners of the above S301 - S302 are the same as those of the above S101 - S102 and will not be elaborated herein.

[0105] S303. Filter the multiple icing features with zero importance according to the feature usage frequency of each icing feature, and / or filter the redundancy of the multiple icing features according to the mutual information between every two icing features to obtain an initial feature set.

[0106] Optionally, filter the multiple icing features with zero importance according to the feature usage frequency of each icing feature to obtain an initial feature set.

[0107] Among them, the feature usage frequency of each icing feature can be used to identify the icing features that contribute nothing to the icing index (such as icing thickness and / or icing growth rate) among the multiple icing features included in the historical icing samples. Thus, these non - contributing icing features are filtered out, and the icing features that contribute to the icing index are obtained. Furthermore, the set composed of the icing features that contribute to the icing index is the initial feature set.

[0108] For example, when the icing index includes icing thickness and icing growth rate, the icing features with a feature usage frequency of 0 for icing thickness and a feature usage frequency of 0 for icing growth rate among the multiple icing features included in the historical icing samples can be filtered out.

[0109] Exemplarily:

[0110]

[0111] Among them, roImportance(a j ) represents the result of zero - importance feature filtering for the j - th icing feature a j , FScore th (a j ) is the feature usage frequency of the j - th icing feature a j for icing thickness, FScore rate (a j ) is the feature usage frequency of the j - th icing feature a j for icing growth rate; otherwise represents the value of the feature usage frequency in other cases except when the feature usage frequency for icing thickness is 0 and the feature usage frequency for icing growth rate is 0. Then, filter out the icing features with ZeroImportance(a j ) = 1 among the multiple icing features included in the historical icing samples.

[0112] Optionally, filter the redundancy of the multiple historical icing samples according to the mutual information between every two icing features to obtain an initial feature set.

[0113] Among them, for every two icing features 1 and 2, the mutual information between the two icing features 1 and 2 can be calculated according to the value probabilities of each sample value of the icing feature 1, the value probabilities of each sample value of the icing feature 2, and the value probabilities of the sample values of the icing feature 1 and the icing feature 2 in the same historical icing sample in the historical icing samples.

[0114] Exemplarily,

[0115] where MI(a i , a j ) is the mutual information between the i-th icing feature a i and the j-th icing feature a j , p(x) is the probability that the sample value of the icing feature a i is x (i.e., the sample proportion of the sample value of the i-th icing feature a i being x in all historical icing samples), p(y) is the probability that the sample value of the j-th icing feature a j is y (i.e., the sample proportion of the sample value of the j-th icing feature a j being y in all historical icing samples), and p(x, y) is the probability that the sample value of the i-th icing feature a i is x and the sample value of the j-th icing feature a j is y (i.e., the sample proportion of the sample value of the i-th icing feature a i being x and the sample value of the j-th icing feature a j being y in all historical icing samples).

[0116] Furthermore, when the above mutual information exceeds a set threshold, it is considered that there is redundancy between the two icing features 1 and 2. Thus, the icing features with mutual information exceeding the set threshold among the multiple icing features included in the historical icing samples are filtered.

[0117] Exemplarily,

[0118] where Redundant(a i , a j ) represents the mutual information filtering result of the i-th icing feature a i and the j-th icing feature a j , θ MI is the set threshold; otherwise represents the value situation of other mutual information except before the mutual information is greater than the set threshold. Then, among the multiple icing features included in the historical icing samples, the i-th icing feature a i , a j )>θ MI and the j-th icing feature a i and the j-th icing feature a j are filtered.

[0119] Optionally, according to the feature usage frequency of each icing feature, zero-importance feature filtering is performed on multiple historical icing samples, and according to the mutual information between every two icing features, redundant filtering is performed on multiple historical icing samples to obtain an initial feature set.

[0120] Exemplarily, as described above, the obtained initial feature set is shown as follows:

[0121] A ′ ={a j |a j ∈A and ZeroImportance(a j ) = 0 and not Redundant(a j )}

[0122] wherein, A ′ is the preliminary feature set, and A is the feature set composed of multiple icing features included in the historical icing samples.

[0123] S304. According to the importance index of each icing feature, using an improved sequential floating forward search method, feature search is performed on multiple icing features included in the initial feature set to obtain a key feature set for icing prediction.

[0124] Among them, the specific implementation manner of the above S304 is similar to the specific implementation manner of the above S303, and will not be elaborated here.

[0125] In this embodiment, first, according to the feature usage frequency of each icing feature, and / or, the mutual information between every two icing features, multiple icing features included in the historical icing samples are filtered, reducing the number of icing features for subsequent feature search using the improved sequential floating forward search method. Thus, the computational burden of feature search is reduced, and the efficiency of icing feature selection is improved.

[0126] Based on the above embodiment, in an exemplary embodiment, the importance index for determining the icing features in S102 is further refined. Optionally, as Figure 4 shown, the following steps may be included:

[0127] S401. Construct an initial model for predicting the icing index, and define the objective function of each decision tree of the initial model.

[0128] Among them, the objective function includes the true value of the icing index.

[0129] It is understandable that in order to achieve icing prediction, it is necessary to establish an initial model for predicting the icing index and screen the icing characteristics for model training. Therefore, an initial model for predicting the icing index is first constructed.

[0130] Optionally, initialize the XGBoost (Extreme Gradient Boosting) model as the initial model for predicting the icing index. Among them, the model parameters of the XGBoost model can be set. For example, the maximum depth D of the tree max , the learning rate η, the regularization parameter λ, and the number of iterations K total , where D max refers to the maximum number of edges passed from the root node to the farthest leaf node of the decision tree of the XGBoost model, and K total refers to the number of decision trees of the XGBoost model.

[0131] After constructing the above initial model, the objective function of each decision tree of the initial model can be further defined to find the best splitting point of each node of the decision tree. Among them, the objective function of each decision tree includes the true value of the icing index in the historical icing samples.

[0132] For example, taking the icing index including the icing thickness and the icing growth rate as an example, two XGBoost models can be initialized as the initial models of the thickness prediction model for predicting the icing thickness and the rate prediction model for predicting the icing growth rate respectively.

[0133] Furthermore, taking the prediction of the icing thickness as an example, in the t-th iteration (i.e., the t-th decision tree), the predicted value of the icing thickness is:

[0134] where is the predicted value of the icing thickness of the historical icing sample i by the t-th tree, is the cumulative predicted value of the icing thickness of the previous t - 1 trees, is the cumulative predicted value of the icing thickness of the previous t trees, and is also the predicted value of the icing thickness in the t-th iteration (i.e., the t-th decision tree).

[0135] Correspondingly, taking the prediction of the icing growth rate as an example, in the t-th iteration (i.e., the t-th decision tree), the predicted value of the icing growth rate is:

[0136] where is the predicted value of the icing growth rate of the historical icing sample i by the t-th tree, the cumulative predicted value of the icing growth rate of the previous t - 1 trees, is the cumulative predicted value of the ice accretion growth rate of the first t trees, and also the predicted value of the ice accretion growth rate in the t-th iteration (i.e., the t-th decision tree).

[0137] Furthermore, for ice thickness prediction, the objective function of each decision tree is:

[0138]

[0139] where is the objective function of the t-th decision tree in the ice thickness prediction model for predicting ice thickness, and n is the total number of historical ice accretion samples.

[0140] For ice accretion growth rate prediction, the objective function of each decision tree is:

[0141]

[0142] where is the objective function of the t-th decision tree in the ice accretion growth rate prediction model for predicting ice accretion growth rate.

[0143] where l th and l rate are the loss functions of the ice thickness prediction model and the ice accretion growth rate prediction model respectively, and Ω(f) is the regularization term, expressed as T is the number of leaf nodes of each decision tree, w j is the weight of the j-th leaf node, and γ is the parameter controlling the complexity of the decision tree.

[0144] Perform a second-order Taylor expansion on the objective function. Taking ice thickness as an example:

[0145]

[0146] where and are the first-order and second-order derivatives of the loss function with respect to the current predicted value respectively.

[0147] For ice thickness prediction, use the mean squared error loss function:

[0148] Similarly, for ice accretion growth rate prediction, use the mean squared error loss function:

[0149] S402. For each decision tree of the initial model, according to the objective function of this decision tree, determine the split gain of each sample value of each ice accretion feature as the feature split point, and according to the split gain, determine the split feature information of each node of this decision tree.

[0150] For each decision tree of the initial model, according to the objective function of the decision tree, the classification gain of each sample value of each icing feature as a feature splitting point can be determined, and based on the obtained splitting gain, starting from the root node of the decision tree, the decision tree can be constructed to determine the splitting feature information of each internal node and leaf node of the decision tree. Among them, the internal nodes of each decision tree are the nodes of the decision tree except the leaf nodes, and the above splitting feature information may include: the number of splitting nodes of each decision tree, the splitting feature of each splitting node, the average splitting gain brought by the splitting feature of each splitting node, the number of samples falling on each splitting node, etc.

[0151] S403. Determine the importance index of each icing feature according to the splitting feature information of each node of each decision tree.

[0152] After obtaining the splitting feature information of each node of each decision tree, the importance index of each icing feature can be determined according to the splitting feature information of each node of each decision tree.

[0153] Among them, the above importance index may include the feature usage frequency and the average splitting gain. Optionally, the above importance index may further include the average coverage.

[0154] Exemplarily, the above importance indexes can be calculated as follows.

[0155]

[0156] Among them, FScore(a j ) is the feature usage frequency of the icing feature a j , K total is the total number of decision trees of the initial model, T k is the total number of splitting nodes of the k-th decision tree of the initial model, v(t) is the splitting feature of each node of each decision tree of the initial model, is the indicator function, when the condition is satisfied, is 1, otherwise, is 0.

[0157]

[0158] Among them, AverageGain(a j ) is the average splitting gain of the icing feature a j , and Gain(t) is the splitting gain of the splitting feature of each node of each decision tree of the initial model.

[0159]

[0160] Among them, AverageCover(aj ) is the average coverage of icing feature a j , and Cover(t) is the number of historical icing samples that fall on each node of each decision tree of the initial model.

[0161] In this embodiment, by constructing an initial model for predicting an icing index and an objective function for each decision tree of the initial model, with the aid of the above objective function and the splitting gain of each sample value of each icing feature as a feature splitting point, each decision tree of the initial model can be constructed starting from the root node. Thus, according to the splitting feature information of each node of each constructed decision tree, the importance index of each icing feature is determined. In this way, the importance index of each icing feature can be determined based on the model features of the initial model for predicting the icing index. Thus, the correlation between the importance index of each icing feature and the initial model for predicting the icing index can be improved. Furthermore, the key features finally selected can have a high correlation with the initial model for predicting the icing index, improving the accuracy of the selected key information. Furthermore, the accuracy of the obtained prediction result is improved, effectively enhancing the power grid's ability to cope with icing disasters.

[0162] As described above, when there are multiple icing indices, icing prediction models for predicting each icing index need to be trained separately. Therefore, when selecting key features, the initial models of the icing prediction models for predicting each icing index should be combined.

[0163] Based on this, on the basis of the above embodiments, in an exemplary embodiment, when the icing indices include icing thickness and icing growth rate, the importance index of the icing feature determined in S102 is further refined. Optionally, as Figure 5 shown, the following steps may be included:

[0164] S501, construct a first initial model for predicting icing thickness and a second initial model for predicting icing growth rate.

[0165] S502, define a first objective function for each first decision tree of the first initial model and a second objective function for each second decision tree of the second initial model.

[0166] S503, for each first decision tree, according to the first objective function of the first decision tree, determine the first splitting gain of each sample value of each icing feature as a feature splitting point, and according to the first splitting gain, determine the splitting feature information of each node of the first decision tree.

[0167] S504. For each second decision tree, according to the second objective function of the second decision tree, determine the second splitting gain of each sample value of each icing feature as a feature splitting point, and according to the first splitting gain, determine the splitting feature information of each node of the second decision tree.

[0168] S505. According to the splitting feature information of each node of each first decision tree, determine the first importance index of each icing feature, and according to the splitting feature information of each node of each second decision tree, determine the second importance index of each icing feature.

[0169] Among them, the specific implementation manners of the above S501 - S505 are similar to the specific implementation manners of the above S401 - S403, and will not be elaborated here.

[0170] S506. For each icing feature, use a preset weight to perform a weighted sum of the first importance index and the second importance index of the icing feature to obtain the importance index of the icing feature.

[0171] As mentioned above, for each icing feature, the first importance index and the second importance index of the icing feature can be obtained, and then a preset weight can be used to perform a weighted sum of the first importance index and the second importance index of the icing feature to obtain the importance index of the icing feature.

[0172] That is to say, for the dual - objective prediction of icing thickness and icing growth rate, the first importance index and the second importance index of each icing feature in the above - mentioned first initial model and the above - mentioned second initial model can be calculated respectively. Furthermore, the above - mentioned first importance index and the second importance index are integrated into a comprehensive importance index as the importance index of the icing feature.

[0173] Exemplarily:

[0174] Scorecombined(a j ) = β × FScoreth(a j )+(1 - β)×FScorerate(a j )

[0175] AverageGain combined (a j ) = β × AverageGain th (a j )+(1 - β)×AverageGain rate (a j )

[0176] AverageCover combined (a j) = β × AverageConver th (a j ) + (1 - β) × AverageCover rate (a j )

[0177] Wherein, FScorecombined(a j ) is the feature usage frequency of icing feature a j , FScoreth(a j ) is the first feature usage frequency of icing feature a j in the first initial model, and FScorerate(a j ) is the second feature usage frequency of icing feature a j in the second initial model;

[0178] AverageGain combined (a j ) is the average split gain of icing feature a j , AverageGain th (a j ) is the first average split gain of icing feature a j in the first initial model, and AverageGain rate (a j ) is the second average split gain of icing feature a j in the second initial model;

[0179] AverageCover combined (a j ) is the average coverage of icing feature a j , AverageCover th (a j ) is the first average coverage of icing feature a j in the first initial model, and AverageCover rate (a j ) is the second average coverage of icing feature a j in the second initial model;

[0180] β is a preset weight, and 0 ≤ β ≤ 1.

[0181] In this embodiment, in the case of multiple icing indicators, the importance indicators of each icing feature under the initial models corresponding to different icing indicators can be calculated separately. Furthermore, through preset weights, the importance indicators of each icing feature under the initial models corresponding to different icing indicators can be balanced to obtain the comprehensive importance indicator of each icing feature for subsequent feature screening. Thus, the accuracy of the selected key information can be improved, and further, the accuracy of the obtained prediction result can be improved, effectively enhancing the power grid's ability to cope with icing disasters.

[0182] Based on the above embodiments, in an exemplary embodiment, the splitting feature information of each node of each decision tree determined in S402 is further refined. Optionally, as Figure 6 shown, the following steps may be included:

[0183] S601, for each decision tree of the initial model, traverse each sample value of each icing feature in the historical icing samples located at the current node of this decision tree. When traversing each sample value of each icing feature, according to the objective function of this decision tree, determine the splitting gain of this sample value of this icing feature.

[0184] For each decision tree of the initial model, there is at least one historical icing sample at its current node, waiting to be divided into the next node of the current node. For example, when the current node of this decision tree is the root node, all historical icing samples are located at the root node, waiting to be divided into the next node of the root node.

[0185] Furthermore, for each decision tree of the initial model, the sample values of the same icing feature included in different historical icing samples located at the current node of this decision tree may be the same or different. Thus, for each icing feature, the respective sample values of this icing feature included in the historical icing samples located at the current node of this decision tree can be determined, and then the above-mentioned respective sample values can be traversed. When traversing each sample value, according to the objective function of this decision tree, determine the splitting gain of this sample value of this icing feature.

[0186] Optionally, in an alternative embodiment, before the above S601, the feature selection method for icing prediction may further include, for each decision tree of the initial model, sorting each sample value of each icing feature in the historical icing samples located at the current node of this decision tree in a specified order to obtain the sample value sequence of each icing feature.

[0187] That is, for each decision tree of the initial model, determine the respective sample values of each icing feature included in each historical icing sample located at the current node of the decision tree, and sort the respective sample values of the same icing feature in the specified order. For example, sort the respective sample values of the same icing feature in ascending or descending order to obtain a sequence of sample values for each icing feature.

[0188] Correspondingly, the above S601 may include traversing the sequence of sample values of each icing feature for each decision tree of the initial model.

[0189] S602. For each decision tree of the initial model, determine the target icing feature corresponding to the obtained maximum splitting gain and the target sample value of the target icing feature.

[0190] As described above, for each decision tree of the initial model, it can determine a splitting gain for each sample value of each icing feature in the historical icing samples located at the current node of the decision tree, so as to obtain a plurality of splitting gains. In this way, the maximum splitting gain among the above-mentioned plurality of splitting gains can be determined, and the target icing feature and the target sample value corresponding to the above maximum splitting gain can be determined.

[0191] That is, for each decision tree of the initial model, the above maximum splitting gain is determined for the target icing feature whose sample value is the target sample value.

[0192] S603. For each decision tree of the initial model, divide the historical icing samples located at the current node of the decision tree into the left and right child nodes of the current node of the decision tree according to the target icing feature and the target sample value.

[0193] For each decision tree of the initial model, after obtaining the above target icing feature and target sample value, the historical icing samples located at the current node of the decision tree can be divided into the left and right child nodes of the current node of the decision tree according to the target icing feature and the target sample value.

[0194] That is, for the historical icing samples located at the current node of the decision tree, divide the historical icing samples whose sample values of the target icing feature are greater than the target sample value into the same child node, and divide the historical icing samples whose sample values of the target icing feature are less than the target sample value into another same child node.

[0195] For example, for each historical icing sample located at the current node of the decision tree, in the case where the sample value of the target icing feature included in the historical icing sample is greater than the target sample value, the historical icing sample is divided into the left child node, and in the case where the sample value of the target icing feature included in the historical icing sample is less than the target sample value, the historical icing sample is divided into the right child node.

[0196] For another example, for each historical icing sample located at the current node of the decision tree, in the case where the sample value of the target icing feature included in the historical icing sample is greater than the target sample value, the historical icing sample is divided into the right child node, and in the case where the sample value of the target icing feature included in the historical icing sample is less than the target sample value, the historical icing sample is divided into the left child node.

[0197] In some cases, due to possible situations such as sensor data loss during the icing monitoring process, the above-mentioned target icing feature may not exist in some historical icing samples. Therefore, for these historical icing samples, it is impossible to directly divide these historical icing samples into the left child node or the right child node according to the numerical relationship between the sample value of the target icing feature they include and the above-mentioned target sample value.

[0198] Based on this, in an optional embodiment, before the subsequent step S604, the feature selection method for icing prediction may further include determining the loss value of dividing the feature-missing sample into the left child node and the loss value of dividing the feature-missing sample into the right child node; and dividing the feature-missing sample into the child node with the minimum loss value among the left child node and the right child node. Wherein, the feature-missing sample is a sample in the historical icing samples of the current node of the decision tree that does not have the target icing feature.

[0199] In this embodiment, for each decision tree of the initial model, a sample in the historical icing samples of the current node of the decision tree that does not have the target icing feature can be used as the feature-missing sample. Furthermore, the loss value of dividing the feature-missing sample into the left child node and the loss value of dividing the feature-missing sample into the right child node can be determined. Thus, the minimum loss value among the above two loss values can be determined, and thus, the feature-missing sample can be divided into the child node corresponding to the above minimum loss value.

[0200] S604, for each decision tree of the initial model, in the case where the preset stop condition is not satisfied, the left child node and the right child node are respectively used as the current node, and S601 is returned for execution.

[0201] For each decision tree of the initial model, when the preset stop condition is not met, it can be considered that the decision tree is not constructed yet and node splitting can continue. Thus, the left and right child nodes obtained after the above splitting can be used as the current node respectively, and the above S301 can be returned to recursively repeat the above process for the left and right child nodes until the preset stop condition is met.

[0202] Optionally, the above preset stop condition can be that the decision tree reaches the maximum depth of the tree, or the maximum splitting gain determined most recently is less than the preset gain threshold.

[0203] S605, for each decision tree of the initial model, when the preset stop condition is met, obtain the constructed decision tree and determine the splitting point information of each node of the decision tree.

[0204] As mentioned above, for each decision tree of the initial model, when the preset stop condition is met after multiple recursions, the node splitting of the decision tree can be stopped and the decision tree can be considered constructed. Thus, the constructed decision tree can be obtained and the splitting point information of each node of the decision tree can be determined.

[0205] In this embodiment, for each decision tree of the initial model, by traversing the sample values of each icing feature in the historical icing samples of the current node of the decision tree, the splitting gain of each sample value of each icing feature can be determined. Thus, based on the splitting gain, the construction of each decision tree of the initial model can be completed to determine the splitting point information of each node of the decision tree. Thereby, the accuracy of the key information selected subsequently according to the splitting point information of each node of each decision tree of the initial model is improved, the accuracy of the obtained prediction result is improved, and the ability of the power grid to cope with icing disasters is effectively improved.

[0206] Based on the above embodiments, in an exemplary embodiment, after obtaining the above key feature set, the icing features in the above key feature set are used to train an icing prediction model.

[0207] Optionally, considering that in actual power grid operation and maintenance, it is necessary to accurately predict the icing thickness and also master the icing growth rate, especially to provide timely early warning for the rapid icing growth stage. Therefore, three interrelated icing prediction models can be trained using the icing features in the above key feature set. These include:

[0208] 1. An icing thickness prediction model, constructing an icing thickness prediction model M based on XGBoost th, the model receives the optimized historical icing characteristics as input and outputs the predicted icing thickness value; among them, the optimized historical icing characteristics include the sample values of the icing characteristics in the above key feature set and the true values of the icing thickness.

[0209]

[0210] Among them, represents the i-th icing sample in the key feature set, is the predicted value of the icing thickness by the icing thickness prediction model M th in the t-th iteration, is the icing thickness prediction model M th in which the predicted value of the icing thickness of the i-th icing sample in the key feature set by the t-th tree.

[0211] 2. Icing growth rate prediction model, based on XGBoost to construct the icing growth rate prediction model M rate , the model receives the optimized historical icing characteristics as input and outputs the predicted icing growth rate; among them, the optimized historical icing characteristics include the sample values of the icing characteristics in the above key feature set and the true values of the icing growth rate.

[0212]

[0213] Among them, is the predicted value of the icing growth rate by the icing growth rate prediction model M rate in the t-th iteration, is the icing growth rate prediction model M rate in which the predicted value of the icing growth rate of the i-th icing sample in the key feature set by the t-th tree.

[0214] 3. Icing rapid growth warning model, used to predict whether the icing enters the rapid growth stage.

[0215]

[0216] Among them, indicates whether the i-th icing sample in the key feature set enters the rapid growth stage; is the predicted value of the icing growth rate of the i-th icing sample in the key feature set by the t-th tree in the icing rapid growth warning model.

[0217] Furthermore, for the icing thickness prediction error and the icing growth rate prediction error, both use the root mean square error (RMSE, Root Mean Square Error) and the mean absolute error (MAE, Mean Absolute Error). Specifically:

[0218]

[0219] Among them, RMSE th represents the root mean square error of the ice thickness prediction model M th , MAE th represents the mean absolute error of the ice thickness prediction model M th , RMSE rate represents the root mean square error of the ice growth rate prediction model M rate , MAE rate represents the mean absolute error of the ice growth rate prediction model M rate , and n test is the number of ice-covered samples in the test samples.

[0220] For the performance of rapid ice growth warning, calculate the warning accuracy, precision, recall, and F1 score:

[0221]

[0222] Among them, Accuracy rapid represents the warning accuracy of the rapid ice growth warning model, Precision rapid represents the precision of the rapid ice growth warning model, Recall rapid represents the recall of the rapid ice growth warning model, and F1 rapid represents the F1 score of the rapid ice growth warning model; TP, FP, TN, and FN are the numbers of true positive, false positive, true negative, and false negative samples of the rapid ice growth warning model, respectively.

[0223] In addition, for each ice-covered feature in the finally obtained key feature set, calculate its comprehensive importance in multi-objective prediction:

[0224]

[0225] Among them, Importance(a j ) is the comprehensive importance of the i-th ice-covered sample in the key feature set.

[0226] Based on the above embodiments, in an exemplary embodiment, as Figure 7 shown, the feature selection method for ice-covered prediction may include the following steps:

[0227] S701. Obtain a plurality of historical ice-covered samples.

[0228] S702. Construct an initial model for predicting ice-covered indicators, and define the objective function of each decision tree of the initial model.

[0229] S703. For each decision tree of the initial model, sort the sample values of each icing feature in the historical icing samples at the current node of the decision tree in a specified order to obtain a sequence of sample values for each icing feature.

[0230] S704. For each decision tree of the initial model, traverse the sequence of sample values of each icing feature. When traversing each sample value of each icing feature, determine the splitting gain of the sample value of the icing feature according to the objective function of the decision tree.

[0231] S705. For each decision tree of the initial model, determine the target icing feature corresponding to the maximum splitting gain obtained and the target sample value of the target icing feature.

[0232] S706. For each decision tree of the initial model, divide the historical icing samples at the current node of the decision tree into the left child node and the right child node of the current node of the decision tree according to the target icing feature and the target sample value.

[0233] S707. Determine the loss value of dividing the feature missing samples into the left child node and the loss value of dividing the feature missing samples into the right child node, and divide the feature missing samples into the child node with the minimum loss value among the left child node and the right child node.

[0234] S708. For each decision tree of the initial model, in the case of not meeting the preset splitting condition, take the left child node and the right child node as the current node respectively, and return to execute S703.

[0235] S709. For each decision tree of the initial model, in the case of meeting the preset splitting condition, obtain the constructed decision tree and determine the splitting point information of each node of the decision tree.

[0236] S710. Filter out zero-importance features for multiple icing features according to the feature usage frequency of each icing feature, and / or filter out redundant features for multiple icing features according to the mutual information between every two icing features to obtain an initial feature set.

[0237] S711. Sort the multiple icing features included in the initial feature set in descending order of feature usage frequency to obtain a first feature subset, and sort the multiple icing features included in the initial feature set in ascending order of average splitting gain to obtain a second feature subset.

[0238] S712. Construct a target feature set, a weighted evaluation function, and a comprehensive evaluation function that are initially empty sets.

[0239] S713. According to the rankings of each icing feature in the first feature subset and the second feature subset, in the first feature subset, search for the first icing feature whose addition makes the function value of the weighted evaluation function of the target feature set increase and the increase value is the largest, and add the first icing feature to the target feature set.

[0240] S714. Delete the first icing feature in the first feature subset to obtain a new first feature subset, and return to execute S713.

[0241] S715. In the case where there is no first icing feature in the first feature subset, according to the rankings of each icing feature in the target feature set in the first feature subset and the second feature subset, in the target feature set, search for the second icing feature whose deletion makes the function value of the comprehensive evaluation function of the target feature set increase.

[0242] S716. Delete the second icing feature in the target feature set, and return to execute S715.

[0243] S717. Determine that there is no second icing feature in the target feature set, and return to execute the step of searching for the first icing feature in the first feature subset whose addition makes the function value of the weighted evaluation function of the target feature set the largest.

[0244] S718. When the target feature set meets the preset stop condition, determine the current target feature set as the key feature set.

[0245] Exemplarily, the case uses the data of a transmission corridor in a certain area in winter 2022, and deploys an icing thickness and growth rate prediction system based on XGBoost. The system collects the following multi-source data.

[0246] 1. Meteorological data, including temperature, humidity, wind speed, wind direction, precipitation, etc. of 16 meteorological stations along the line, with a sampling frequency of 15 minutes; regional meteorological data and numerical weather prediction products provided by the China Meteorological Administration. 2. Line parameters, including conductor type; parameters such as line height, inclination angle, azimuth angle, etc.; tower distribution and span information. 3. Historical icing records, icing monitoring data from 2018 to 2021, including icing thickness and icing growth rate; detailed records of 30 significant icing events, including 7 severe icing events that led to line tripping. 4. Real-time monitoring data, real-time data of 8 sets of icing online monitoring devices installed at key positions; transmission line load and conductor surface temperature sensor data.

[0247] The first step is feature extraction and construction. Specifically, after preprocessing the original data, an initial feature set is constructed, which contains 167 icing features. Among them, there are 42 direct meteorological features, 78 time-series derived features, 23 line features, 14 geographical features, and 10 combined features. The second step is to calculate the feature importance measure. Two XGBoost models are used for the icing thickness and growth rate prediction tasks respectively. The model parameters are set as the maximum tree depth of 6, the learning rate of 0.05, the number of iterations of 200, and the regularization parameter λ of 1.5. And three importance indicators for each icing feature are calculated: FScore, AverageGain, and AverageCover. Among them, the importance indicators of some icing features are shown in Table 1 below:

[0248] Table 1 Importance indicators of icing features

[0249]

[0250]

[0251] The third step is feature preprocessing. It is detected that there are about 15% missing values in the original data, mainly occurring under extreme weather conditions. After applying the missing value processing mechanism of XGBoost, these samples are effectively utilized, and through zero importance filtering and redundant feature detection, 59 icing features are initially removed, reducing the feature set to 108. The fourth step is ISFFS feature search. An improved sequential floating forward search strategy is applied, and weights for different icing stages are set.

[0252] Among them, in the initial formation stage: γ_initial = 0.6 (more focused on thickness prediction), in the stable growth stage: γ_stable = 0.5 (thickness and rate are equally important), in the rapid growth stage: γ_rapid = 0.3 (more focused on rate prediction). Furthermore, after 12 rounds of iteration, the ISFFS algorithm converges, and finally 23 key features are selected, including: 5 direct meteorological features (current temperature, relative humidity, wind speed, air pressure, precipitation); 7 time-series derived features (temperature change rate, temperature 3 hours ago, etc.); 4 line features (conductor height, inclination, etc.); 3 geographical features (altitude, slope aspect, etc.); and 4 combined features (temperature-humidity-wind composite index, etc.).

[0253] The fifth step: Performance evaluation: Remove the features in the original data that do not belong to the above key features, and divide the original data after feature removal into a training set and a test set according to the ratio of 80%:20%. Evaluate the model performance of the icing thickness prediction model and the icing growth rate prediction model trained using the original data after feature removal on the test set. Among them,

[0254] The model performance of the ice thickness prediction model includes: RMSE is 2.34 mm (better than 3.87 mm of the baseline method), and MAE is 1.78 mm (better than 2.92 mm of the baseline method). The model performance of the ice growth rate prediction model includes: RMSE is 0.31 mm / h (better than 0.52 mm / h of the baseline method), and MAE is 0.23 mm / h (better than 0.41 mm / h of the baseline method).

[0255] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least some of the steps or stages in other steps or other steps.

[0256] Based on the same inventive concept, an embodiment of the present application further provides an ice coverage prediction feature selection device for implementing the ice coverage prediction feature selection method described above. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the ice coverage prediction feature selection device provided below can refer to the limitations on the ice coverage prediction feature selection method in the above text, and will not be repeated here.

[0257] In an exemplary embodiment, as Figure 8 shown, an ice coverage prediction feature selection device is provided, including: a sample acquisition module 810, an index determination module 820, and a feature search module 830, where:

[0258] The sample acquisition module 810 acquires a plurality of historical ice coverage samples; wherein, the historical ice coverage samples include the true values of ice coverage indicators and the sample values of a plurality of ice coverage features, and the ice coverage indicators include ice thickness and / or ice growth rate;

[0259] The index determination module 820 is configured to determine the importance index of each ice coverage feature according to the true value and the feature value of the ice coverage feature in the same historical ice coverage sample; wherein, the importance index includes at least the feature usage frequency and the average split gain;

[0260] A feature search module 830 is configured to perform feature search on multiple icing features by using an improved sequential floating forward search method according to the importance index of each icing feature, so as to obtain a key feature set for icing prediction.

[0261] In an exemplary embodiment, the index determination module 820 includes: a model construction unit configured to construct an initial model for predicting an icing index and define an objective function for each decision tree of the initial model; wherein the objective function includes the true value of the icing index; an information determination unit configured to, for each decision tree of the initial model, determine the splitting gain of each sample value of each icing feature as a feature splitting point according to the objective function of the decision tree, and determine the splitting feature information of each node of the decision tree according to the splitting gain; an index determination unit configured to determine the importance index of each icing feature according to the splitting feature information of each node of each decision tree.

[0262] In an exemplary embodiment, the information determination unit includes: a gain determination subunit configured to, for each decision tree of the initial model, traverse each sample value of each icing feature in the historical icing samples located at the current node of the decision tree, and determine the splitting gain of the sample value of the icing feature according to the objective function of the decision tree when traversing each sample value of each icing feature; a gain selection subunit configured to, for each decision tree of the initial model, determine the target icing feature corresponding to the obtained maximum splitting gain and the target sample value of the target icing feature; a sample division subunit configured to, for each decision tree of the initial model, divide the historical icing samples located at the current node of the decision tree into the left child node and the right child node of the current node of the decision tree according to the target icing feature and the target sample value; a node update subunit configured to, for each decision tree of the initial model, when the preset splitting condition is not satisfied, use the left child node and the right child node as the current node respectively, and trigger the gain determination subunit; an information determination subunit configured to, for each decision tree of the initial model, when the preset splitting condition is satisfied, obtain the constructed decision tree and determine the splitting point information of each node of the decision tree.

[0263] In an exemplary embodiment, the feature selection device for icing prediction further includes: a sample value sorting module configured to, for each decision tree of the initial model, sort each sample value of each icing feature in the historical icing samples located at the current node of the decision tree in a specified order to obtain a sample value sequence of each icing feature; the gain determination subunit is specifically configured to: for each decision tree of the initial model, traverse the sample value sequence of each icing feature.

[0264] In an exemplary embodiment, the feature selection device for icing prediction further includes: a loss value calculation module, configured to determine the loss value of dividing the feature-missing samples into the left child node and the loss value of dividing the feature-missing samples into the right child node; and divide the feature-missing samples into the child node with the minimum loss value among the left child node and the right child node; wherein the feature-missing samples are the samples in the historical icing samples of the current node of the decision tree that do not have the target icing feature.

[0265] In an exemplary embodiment, the icing metrics include icing thickness and icing growth rate; the model construction unit is specifically configured to construct a first initial model for predicting the icing thickness and a second initial model for predicting the icing growth rate; define the first objective function of each first decision tree of the first initial model and the second objective function of each second decision tree of the second initial model; the information determination unit is specifically configured to, for each first decision tree, according to the first objective function of the first decision tree, determine the first splitting gain of each sample value of each icing feature as a feature splitting point, and determine the splitting feature information of each node of the first decision tree according to the first splitting gain; for each second decision tree, according to the second objective function of the second decision tree, determine the second splitting gain of each sample value of each icing feature as a feature splitting point, and determine the splitting feature information of each node of the second decision tree according to the first splitting gain; the metric determination unit is specifically configured to determine the first importance metric of each icing feature according to the splitting feature information of each node of each first decision tree, and determine the second importance metric of each icing feature according to the splitting feature information of each node of each second decision tree; for each icing feature, use a preset weight to perform a weighted sum of the first importance metric and the second importance metric of the icing feature to obtain the importance metric of the icing feature.

[0266] In an exemplary embodiment, the feature search module 830 is specifically configured to: sort multiple icing features in descending order of feature usage frequency to obtain a first feature subset, and sort multiple icing features in ascending order of average split gain to obtain a second feature subset; construct a target feature set, a weighted evaluation function, and a comprehensive evaluation function that are initially empty sets; according to the sorting of each icing feature in the first feature subset and the second feature subset, in the first feature subset, search for a first icing feature that, after being added, increases the function value of the weighted evaluation function of the target feature set and has the largest increase value, and add the first icing to the target feature set; delete the first icing feature in the first feature subset to obtain a new first feature subset, and return to execute the step of searching for the first icing feature in the first feature subset that, after being added, maximizes the function value of the weighted evaluation function of the target feature set; in the case where there is no first icing feature in the first feature subset, according to the sorting of each icing feature in the target feature set in the first feature subset and the second feature subset, search for a second icing feature in the target feature set that, after being deleted, increases the function value of the comprehensive evaluation function of the target feature set; delete the second icing feature in the target feature set, and return to execute the step of searching for the second icing feature in the target feature set that, after being deleted, increases the function value of the comprehensive evaluation function of the target feature set; determine that there is no second icing feature in the target feature set, and return to execute the step of searching for the first icing feature in the first feature subset that, after being added, maximizes the function value of the weighted evaluation function of the target feature set; in the case where the target feature set meets the preset stop condition, determine the current target feature set as the key feature set.

[0267] In an exemplary embodiment, the feature selection device for icing prediction further includes: a feature filtering module, configured to perform zero-importance feature filtering on multiple icing features according to the feature usage frequency of each icing feature, and / or perform redundancy filtering on multiple icing features according to the mutual information between every two icing features to obtain an initial feature set.

[0268] The feature search module 830 is specifically configured to: perform feature search on multiple icing features included in the initial feature set by using an improved sequential floating forward search method according to the importance index of each icing feature to obtain a key feature set for icing prediction.

[0269] Each module in the above feature selection device for icing prediction can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in or independent of a processor in a computer device in the form of hardware, or stored in a memory in the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above respective modules.

[0270] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structural diagram may be as shown in Figure 9 . The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store data such as historical icing samples. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a method for a feature selection device for icing prediction.

[0271] Those skilled in the art can understand that Figure 9 the structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0272] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps of the above method embodiments are implemented.

[0273] In an embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, the steps of the above method embodiments are implemented.

[0274] In an embodiment, a computer program product is provided, including a computer program. When the computer program is executed by the processor, the steps of the above method embodiments are implemented.

[0275] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, artificial intelligence (AI) processors, etc., without limitation.

[0276] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in the present application.

[0277] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A feature selection method for icing prediction, characterized in that The method includes: Obtaining a plurality of historical icing samples; wherein, the historical icing samples include the true values of icing indicators and the sample values of a plurality of icing characteristics, and the icing indicators include icing thickness and / or icing growth rate; For each icing characteristic, according to the true value in the same historical icing sample and the characteristic value of this icing characteristic, determining the importance index of this icing characteristic; wherein, the importance index at least includes feature usage frequency and average splitting gain; According to the importance index of each icing characteristic, using an improved sequential floating forward search method to perform feature search on the plurality of icing characteristics to obtain a key feature set for icing prediction.

2. The method according to claim 1, characterized in that, The step of, for each icing characteristic, according to the true value in the same historical icing sample and the characteristic value of this icing characteristic, determining the importance index of this icing characteristic includes: Constructing an initial model for predicting the icing indicator and defining the objective function of each decision tree of the initial model; wherein, the objective function includes the true value of the icing indicator; For each decision tree of the initial model, according to the objective function of this decision tree, determining the splitting gain of each sample value of each icing characteristic as a feature splitting point, and according to the splitting gain, determining the splitting feature information of each node of this decision tree; According to the splitting feature information of each node of each decision tree, determining the importance index of each icing characteristic.

3. The method according to claim 2, wherein The step of, for each decision tree of the initial model, according to the objective function of this decision tree, determining the splitting gain of each sample value of each icing characteristic as a feature splitting point, and according to the splitting gain, determining the splitting point information of each node of this decision tree includes: For each decision tree of the initial model, traversing each sample value of each icing characteristic in the historical icing samples located at the current node of this decision tree, and when traversing each sample value of each icing characteristic, determining the splitting gain of this sample value of this icing characteristic according to the objective function of this decision tree; For each decision tree of the initial model, determining the target icing characteristic corresponding to the obtained maximum splitting gain and the target sample value of the target icing characteristic; For each decision tree of the initial model, according to the target icing characteristic and the target sample value, dividing the historical icing samples located at the current node of this decision tree into the left child node and the right child node of the current node of this decision tree; For each decision tree of the initial model, when the preset splitting condition is not met, taking the left child node and the right child node as the current nodes respectively, and returning to execute the step of traversing each sample value of each icing characteristic in the historical icing samples located at the current node of this decision tree; For each decision tree of the initial model, when the preset splitting condition is met, obtaining the constructed decision tree and determining the splitting point information of each node of this decision tree.

4. The method according to claim 3, characterized in that, Before the step of, for each decision tree of the initial model, traversing each sample value of each icing characteristic in the historical icing samples located at the current node of this decision tree, the method further includes: For each decision tree of the initial model, in a specified order, sort the sample values of each icing feature in the historical icing samples at the current node of the decision tree to obtain a sequence of sample values for each icing feature; The traversal of the sample values of each icing feature in the historical icing samples at the current node of each decision tree of the initial model includes: For each decision tree of the initial model, traverse the sequence of sample values of each icing feature.

5. The method according to claim 3, wherein Before taking the left child node and the right child node as the current node respectively for each decision tree of the initial model when the preset stop condition is not met, the method further includes: Determine the loss value of dividing the feature - missing samples into the left child node and the loss value of dividing the feature - missing samples into the right child node; wherein, the feature - missing samples are the samples in the historical icing samples at the current node of the decision tree that do not have the target icing feature; Divide the feature - missing samples into the child node with the minimum loss value among the left child node and the right child node.

6. The method according to claim 2, characterized in that, The icing metrics include the icing thickness and the icing growth rate; Building an initial model for predicting the icing metrics and defining the objective function of each decision tree of the initial model includes: Build a first initial model for predicting the icing thickness and a second initial model for predicting the icing growth rate; Define the first objective function of each first decision tree of the first initial model and the second objective function of each second decision tree of the second initial model; For each decision tree of the initial model, according to the objective function of the decision tree, determine the split gain of each sample value of each icing feature as a feature split point, and according to the split gain, determine the split point information of each node of the decision tree, including: For each first decision tree, according to the first objective function of the first decision tree, determine the first split gain of each sample value of each icing feature as a feature split point, and according to the first split gain, determine the split feature information of each node of the first decision tree; For each second decision tree, according to the second objective function of the second decision tree, determine the second split gain of each sample value of each icing feature as a feature split point, and according to the first split gain, determine the split feature information of each node of the second decision tree; Determining the importance index of each icing feature according to the split feature information of each node of each decision tree includes: According to the split feature information of each node of each first decision tree, determine the first importance index of each icing feature, and according to the split feature information of each node of each second decision tree, determine the second importance index of each icing feature; For each icing feature, use a preset weight to perform a weighted sum of the first importance index and the second importance index of the icing feature to obtain the importance index of the icing feature.

7. The method according to any one of claims 1-6, characterized in that According to the importance index of each icing feature, using an improved sequential floating forward search method, feature search is performed on the multiple icing features to obtain a key feature set for icing prediction, including: Sort the multiple icing features in descending order of the feature usage frequency to obtain a first feature subset, and sort the multiple icing features in ascending order of the average split gain to obtain a second feature subset; Construct a target feature set, a weighted evaluation function, and a comprehensive evaluation function that are initially empty sets; According to the sorting of each icing feature in the first feature subset and the second feature subset, in the first feature subset, search for the first icing feature that makes the function value of the weighted evaluation function of the target feature set increase and the increase value is the largest after addition, and add the first icing to the target feature set; Delete the first icing feature in the first feature subset to obtain a new first feature subset, and return to execute the step of searching for the first icing feature that makes the function value of the weighted evaluation function of the target feature set the largest after addition in the first feature subset; In the case that the first icing feature does not exist in the first feature subset, according to the sorting of each icing feature in the target feature set in the first feature subset and the second feature subset, search for the second icing feature in the target feature set that makes the function value of the comprehensive evaluation function of the target feature set increase after deletion; Delete the second icing feature in the target feature set, and return to execute the step of searching for the second icing feature that makes the function value of the comprehensive evaluation function of the target feature set increase after deletion in the target feature set; Determine that the second icing feature does not exist in the target feature set, and return to execute the step of searching for the first icing feature that makes the function value of the weighted evaluation function of the target feature set the largest after addition in the first feature subset; In the case that the target feature set meets the preset stop condition, determine the current target feature set as the key feature set.

8. The method according to any one of claims 1 to 6, characterized in that, Before performing the feature search on the multiple icing features according to the importance index of each icing feature using an improved sequential floating forward search method to obtain a key feature set for icing prediction, the method further includes: Perform zero-importance feature filtering on the multiple icing features according to the feature usage frequency of each icing feature, and / or perform redundant filtering on the multiple icing features according to the mutual information between every two icing features to obtain an initial feature set; The feature search on the multiple icing features according to the importance index of each icing feature using an improved sequential floating forward search method to obtain a key feature set for icing prediction includes: According to the importance index of each icing feature, use an improved sequential floating forward search method to perform feature search on the multiple icing features included in the initial feature set to obtain a key feature set for icing prediction.

9. A feature selection device for icing prediction, characterized in that The device includes: A sample acquisition module that acquires a plurality of historical icing samples; wherein, the historical icing samples include the true values of icing indicators and the sample values of a plurality of icing characteristics, and the icing indicators include icing thickness and / or icing growth rate; An indicator determination module for determining, for each icing characteristic, an importance indicator of the icing characteristic according to the true value and the characteristic value of the icing characteristic in the same historical icing sample; wherein, the importance indicator includes at least the feature usage frequency and the average splitting gain; A feature search module for performing feature search on the plurality of icing characteristics by using an improved sequential floating forward search method according to the importance indicators of each icing characteristic, so as to obtain a key feature set for icing prediction.

10. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method described in any one of claims 1 to 8 are implemented.

Images