Method for determining engineering quantity interval of technical improvement project of power grid

Abstract

The invention provides a power grid technical improvement project work amount interval determination method comprising the following steps: obtaining construction cost data of a power grid technical improvement project, the construction cost data comprising historical work amount data, regional economic data, market price fluctuation data, construction environment data, and meteorological condition data; preprocessing the cost data to obtain standard data; performing feature extraction on the standard data to obtain a high-dimensional feature vector; inputting the high-dimensional feature vector into a neural network model to obtain an engineering quantity interval prediction result output by the neural network model; the neural network model is a multi-task deep neural network model obtained by training according to training data, and the training data comprises historical cost data. According to the method, feature differences of geographical environments, climate conditions and engineering types of different regions are fully considered in the process of calculating the engineering amount interval. Through data fusion and feature extraction, the problem of inaccurate calculation of a project amount interval of a power grid technical improvement project is solved.

Description

Technical Field

[0001] The present application relates to the technical field of methods for predicting engineering quantity intervals for technical transformation projects, and specifically to a method for determining engineering quantity intervals for power grid technical transformation projects. Background Art

[0002] With the rapid development of social economy, the demand for power grid infrastructure construction and transformation is increasing, and power grid technical transformation projects have gradually become an important means for power grid companies to improve power supply reliability and optimize resource allocation. Power grid technical transformation projects usually include multiple fields such as line transformation, substation upgrades, and distribution network expansion, and their scale and complexity are increasing year by year. However, due to significant differences in geographical environment, construction conditions, project types and other factors in various regions, the engineering quantity assessment of power grid technical transformation projects faces many problems. Scientifically and accurately assessing the engineering quantity range of power grid technical transformation projects has become a key issue for power grid companies to carry out project planning and resource allocation.

[0003] Traditional methods for evaluating the quantity of power grid projects mainly rely on engineers' experience and historical project data. Although these methods can provide a basis for estimation to a certain extent, they fail to fully consider the environmental conditions in different regions and the differences in construction difficulty of different types of projects, resulting in large deviations in the evaluation results. Especially in areas with complex terrain and changeable climate, traditional methods are difficult to accurately predict the quantity of projects, which affects the resource allocation and construction progress of the project. In addition, as power grid technical transformation projects gradually develop towards refined management, how to achieve standardization, normalization and scientificization of project quantity evaluation is an important technical problem faced by power grid companies.

[0004] In the existing engineering quantity assessment methods, the usual practice is to make simple ratio adjustments based on historical averages. This method cannot accurately reflect the actual impact of geological conditions, climate environments, etc. in different regions on engineering quantities. For example, when carrying out line reconstruction in mountainous and hilly areas, the amount of engineering may increase significantly due to steep terrain and difficult construction, while it is relatively less in plain areas. In addition, different types of technical transformation projects, such as transmission line upgrades and substation expansions, also have obvious differences in their engineering quantity composition and influencing factors. If these factors are not refined, the final assessment results may deviate significantly from the actual situation. Summary of the invention

[0005] The present application provides a method for determining an engineering quantity interval for a power grid technical transformation project, so as to solve the problem of inaccurate engineering quantity interval calculation for the power grid technical transformation project.

[0006] The present application provides a method for determining an engineering quantity interval of a power grid technical transformation project, comprising:

[0007] Obtaining cost data for power grid technical transformation projects, the cost data including historical engineering quantity data, regional economic data, market price fluctuation data, construction environment data, and meteorological condition data;

[0008] Preprocessing the cost data to obtain standard data;

[0009] Performing feature extraction on the standard data to obtain a high-dimensional feature vector;

[0010] The high-dimensional feature vector is input into a neural network model to obtain an engineering quantity interval prediction result output by the neural network model; the neural network model is a multi-task deep neural network model trained according to training data, and the training data includes historical cost data.

[0011] Optionally, performing preprocessing on the cost data to obtain standard data includes:

[0012] Performing data cleaning on the construction cost data to obtain first standard data;

[0013] Normalization and linear discriminant analysis are performed on the first standard data to obtain the standard data.

[0014] Optionally, performing data cleaning on the construction cost data to obtain the first standard data includes:

[0015] Performing mean filling on the construction cost data to fill missing values ​​of the construction cost data;

[0016] executing an adaptive clustering algorithm on the cost data to delete outliers in the cost data;

[0017] The cost data is traversed to delete repeated values ​​in the cost data to obtain first standard data.

[0018] Optionally, performing normalization and linear discriminant analysis on the first standard data to obtain the standard data includes:

[0019] Performing minimum-maximum value normalization on the first standard data to obtain normalized data;

[0020] Calculating an intra-class hash matrix and an inter-class hash matrix of the normalized data;

[0021] Solving the eigenvalue and eigenvector of the product of the inverse matrix of the intra-class hash matrix and the inter-class hash matrix to obtain a projection direction;

[0022] The normalized data is projected onto a projection direction to obtain standard data.

[0023] Optionally, performing feature extraction on the standard data to obtain a high-dimensional feature vector includes:

[0024] Constructing a multi-dimensional feature map according to the standard data;

[0025] Inputting the multi-dimensional feature map into a graph convolutional network for fusion to obtain fused data;

[0026] The fused data is input into a convolutional neural network for feature extraction to generate a high-dimensional feature vector; the convolutional neural network is provided with an attention mechanism.

[0027] Optionally, the method further includes:

[0028] Obtaining a preset amount of the historical cost data;

[0029] Randomly extracting a number of the historical cost data as a training set;

[0030] Constructing a multi-task deep neural network model; the multi-task deep neural network model includes an input layer, a shared layer, a sub-network layer, a shared output layer and a sub-network output layer; the input layer is used to receive a high-dimensional feature vector, and the shared layer extracts shared underlying features in the high-dimensional feature vector through a multi-layer fully connected network;

[0031] The historical cost data is input into the multi-task deep neural network model, and forward propagation and back propagation are performed according to the output results of the multi-task deep neural network model to obtain the multi-task deep neural network model that has completed training; the forward propagation is to calculate the predicted output and loss value of the multi-task deep neural network model. The back propagation is to calculate the gradient according to the loss value and update the model parameters by the gradient descent method.

[0032] Optionally, building a multi-task deep neural network model includes:

[0033] Setting the initial network weights of the multi-task deep network model, wherein the initial network weights are network weights of a pre-trained model related to the engineering quantity prediction task;

[0034] performing fine-tuning on the initial network weights;

[0035] The multi-task deep neural network model is optimized by a hybrid loss function.

[0036] Optionally, also include:

[0037] According to the Bayesian optimization algorithm, the hyperparameters of the multi-task deep neural network model are adjusted; the hyperparameters include learning rate, number of network layers, and number of nodes;

[0038] According to the dynamic adjustment algorithm, the engineering quantity interval results are generated based on the engineering quantity interval prediction results.

[0039] Optionally, adjusting the hyperparameters of the multi-task deep neural network model according to a Bayesian optimization algorithm includes:

[0040] Building agent models;

[0041] Obtaining hyperparameters of performance data of the multi-task deep neural network to generate preset hyperparameters according to the proxy model;

[0042] Calculate the expected improvement value according to the preset hyperparameters;

[0043] The hyperparameter is updated according to the expected improvement value.

[0044] Optionally, generating the engineering quantity interval result based on the engineering quantity interval prediction result according to the dynamic adjustment algorithm includes:

[0045] Calculating the confidence interval of the preset result of the engineering quantity interval;

[0046] Based on the confidence interval, a risk factor is added to adjust the preset result of the engineering quantity interval to generate an engineering quantity interval result.

[0047] It can be seen from the above technical solutions that the present application provides a method for determining the engineering quantity interval of a power grid technical transformation project, including: obtaining the cost data of the power grid technical transformation project, the cost data including historical engineering quantity data, regional economic data, market price fluctuation data, construction environment data, and meteorological condition data; performing preprocessing on the cost data to obtain standard data; performing feature extraction on the standard data to obtain a high-dimensional feature vector; inputting the high-dimensional feature vector into a neural network model to obtain the engineering quantity interval prediction result output by the neural network model; the neural network model is a multi-task deep neural network model obtained by training according to training data, and the training data includes historical cost data. In the process of calculating the engineering quantity interval, the present application fully considers the characteristic differences in geographical environment, climatic conditions and engineering types in different regions. Through data fusion and feature extraction, the problem of inaccurate calculation of engineering quantity intervals for power grid technical transformation projects is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Figure 1 A flow chart of a method for determining an engineering quantity interval for a power grid technical transformation project provided in an embodiment of the present application;

[0050] Figure 2 A flow chart of a method for preprocessing cost data provided in an embodiment of the present application;

[0051] Figure 3 Provide a flow chart of a standard data feature extraction method for the embodiment of the present application;

[0052] Figure 4 A multi-task deep neural network model training flow chart provided for an embodiment of the present application;

[0053] Figure 5 A flowchart of a method for constructing a multi-task deep neural network model provided in an embodiment of the present application;

[0054] Figure 6 A flow chart of a method for determining an engineering quantity interval provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The following embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following embodiments do not represent all implementations consistent with the present application. They are only examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0056] With the rapid development of social economy, the demand for power grid infrastructure construction and transformation is increasing, and power grid technical transformation projects have gradually become an important means for power grid companies to improve power supply reliability and optimize resource allocation. Power grid technical transformation projects usually include multiple fields such as line transformation, substation upgrades, and distribution network expansion, and their scale and complexity are increasing year by year. However, due to significant differences in geographical environment, construction conditions, project types and other factors in various regions, the engineering quantity assessment of power grid technical transformation projects faces many problems. Scientifically and accurately assessing the engineering quantity range of power grid technical transformation projects has become a key issue for power grid companies to carry out project planning and resource allocation.

[0057] In some embodiments, the engineering volume range of the power grid technical transformation project is obtained by calculating the historical average and then making a simple ratio adjustment. However, the above method cannot accurately reflect the actual impact of geological conditions, climate environment, etc. in different regions on the engineering volume. For example, in mountainous and hilly areas, the engineering volume may increase significantly due to steep terrain and difficult construction, while it is relatively less in plain areas. In addition, different types of technical transformation projects, such as transmission line upgrades and substation expansions, also have obvious differences in the composition and influencing factors of their engineering volumes. If these factors are not refined, the final evaluation results may deviate significantly from the actual situation.

[0058] Therefore, in order to solve the problem of inaccurate calculation of engineering quantity intervals for power grid technical transformation projects, the present application embodiment provides a method for determining engineering quantity intervals for power grid technical transformation projects, such as Figure 1 As shown, including:

[0059] S100: Obtaining cost data of power grid technical transformation projects;

[0060] Among them, construction cost data includes historical engineering quantity data, regional economic data, market price fluctuation data, construction environment data, and meteorological conditions data.

[0061] Specifically, the historical engineering quantity data includes the historical engineering quantities of various power grid technical transformation projects, which can be classified by region, project, and type.

[0062] Regional economic data include GDP, per capita income, consumption level and other indicators that reflect the regional economic level.

[0063] Market fluctuation data include data such as the prices of materials needed in power grid technical transformation projects and local labor costs.

[0064] Construction environment data includes terrain, soil type, traffic conditions and other data that affect construction costs.

[0065] Meteorological condition data include average temperature, rainfall, wind speed and other meteorological data that affect construction difficulty and cost.

[0066] By acquiring and integrating data from different sources, taking into account the differences in geographical environment, climatic conditions and engineering types in different regions, and through data fusion and feature extraction, the evaluation error problem caused by ignoring regional and type differences in traditional methods is solved, providing a scientific basis for the planning and resource allocation of power grid technical transformation projects.

[0067] S200: Preprocessing the cost data to obtain standard data.

[0068] After obtaining the cost data, it is necessary to preprocess the cost data to ensure data quality and the reliability of subsequent analysis. Figure 2 As shown, the steps of preprocessing the cost data to obtain standard data include:

[0069] S210: performing data cleaning on the cost data to obtain first standard data;

[0070] Among them, data cleaning of cost data includes:

[0071] Perform mean imputation on the cost data to fill in the missing values ​​of the cost data.

[0072] An adaptive clustering algorithm is performed on the cost data to remove outliers in the cost data.

[0073] The cost data is traversed to delete duplicate values ​​in the cost data to obtain first standard data.

[0074] Specifically, since there may be some missing data in the acquired cost data, it is necessary to process the missing values. Common methods for processing missing values ​​include deleting records with missing values, filling missing values ​​with mean / median / mode, or using more complex interpolation methods. In the embodiment of the present application, when the missing value is numerical data, the mean filling method is used to process the missing value, wherein the formula for mean filling is:

[0075]

[0076] Among them, x filled is the missing value after filling, m is the number of non-missing values, x i is known data.

[0077] After the missing values ​​are processed, an adaptive algorithm is needed to detect and delete outliers in the cost data. For example, the DBSCAN algorithm is a density-based clustering algorithm that can effectively identify outliers.

[0078] Among them, density reachability is defined in the DBSCAN algorithm, that is, if point P includes at least MinPts (minimum number of points) points within ∈ (neighborhood radius), then point P is called a core point, and the outlier value is defined as if point P is neither a core point nor within the ∈ neighborhood of any core point, then it is called an outlier point.

[0079] Therefore, when using it, the parameters ∈ and MinPts can be initialized. For each point in the cost data, the number of points in its neighborhood is calculated. If the number of points in the neighborhood is ≥ MinPts, the point is marked as a core point; otherwise, it is marked as a boundary point or a noise point. It can be understood that the boundary point or the noise point is an outlier. After obtaining the outlier, the outlier can be deleted.

[0080] Finally, it is also necessary to deal with the duplicate values ​​in the cost data. You can do this by traversing the data, filtering out the duplicate values, deleting the duplicate values ​​and retaining one data value.

[0081] After the missing values, abnormal values ​​and duplicate values ​​in the cost data are processed, the first standard data can be obtained. However, due to the different magnitude ranges and dimensions of different data, normalization and linear discriminant analysis are required to obtain standard data. Among them, performing normalization and linear discriminant analysis on the first standard data includes:

[0082] S220: Perform minimum-maximum value normalization on the first standard data to obtain normalized data.

[0083] S230: Calculate the intra-class hash matrix and the inter-class hash matrix of the normalized data.

[0084] S240: Solve the eigenvalues ​​and eigenvectors of the product of the inverse matrix of the intra-class hash matrix and the inter-class hash matrix to obtain the projection direction.

[0085] S250: Projecting the normalized data onto the projection direction to obtain standard data.

[0086] Specifically, the first standard data is substituted into the minimum-maximum value normalization formula to obtain normalized data, wherein the minimum-maximum value normalization formula is:

[0087]

[0088] Among them, x is the original data value, x min and x max are the minimum and maximum values ​​of the data, respectively, * is the normalized data value.

[0089] After obtaining the normalized data through calculation, it is also necessary to perform linear discriminant analysis on the normalized data. For example, using the LDA supervised learning method, the normalized data can be classified and reduced in dimension. The normalized data can be projected into a low-dimensional space through linear transformation, and data points of different categories can be distinguished at the same time, thereby reducing the computational complexity.

[0090] The formula for calculating the intra-class scatter matrix and the inter-class scatter matrix is:

[0091]

[0092] Among them, k is the number of categories, x is the feature vector of the sample, μ i is the mean vector of the i-th class samples, C i is the category label of the jth sample; S W Represents a d×d matrix, where d is the number of features, which represents the degree of difference between samples within the same category.

[0093]

[0094] Among them, k is the number of categories, n i The number of samples in the i-th class, μ i is the mean vector of the i-th sample, μ represents the overall mean vector of all samples, S B represents a d×d matrix, where d is the number of features, which represents the degree of difference in sample means between different categories.

[0095] Solve the inverse matrix of the within-class hash matrix and the inter-class hash matrix S B The product of the eigenvalues ​​and eigenvectors gives the projection direction.

[0096] Among them, the optimal projection direction can be found by maximizing the objective function, and the maximum objective function is:

[0097]

[0098] By solving the generalized eigenvalue problem The eigenvector of can be used to find the projection direction ω. Finally, the normalized data is projected onto the projection direction ω to obtain the standard data.

[0099] Through the above steps and formulas, the process of data cleaning, normalization, feature selection and dimensionality reduction is completed, which can ensure the high quality and efficiency of the data and provide a reliable foundation for subsequent deep learning model training.

[0100] S300: Perform feature extraction on the standard data to obtain a high-dimensional feature vector.

[0101] After obtaining the standard data, it is also necessary to extract features from the standard data and integrate the data from different data sources to obtain the relationship and interaction between the data from different data sources. Figure 3 As shown, feature extraction is performed on the standard data to obtain a high-dimensional feature vector, which specifically includes:

[0102] S310: Construct a multi-dimensional feature map based on standard data.

[0103] S320: Input the multi-dimensional feature map into the graph convolutional network for fusion to obtain fused data.

[0104] S330: Input the fused data into a convolutional neural network for feature extraction to generate a high-dimensional feature vector.

[0105] Among them, the multidimensional feature map represents the relationship and mutual influence between various data. The nodes in the multidimensional feature map represent various data sources. For example, historical engineering cost data, regional economic data, and market fluctuation data can be represented as nodes, and the edges between nodes represent the relationship or similarity between various data sources. For example, the economic data of adjacent regions can be represented by edges to represent their geographical proximity.

[0106] Specifically, suppose that the multidimensional feature graph G = (V, E) is composed of a node set V and an edge set E, where node vi∈V represents the i-th feature, and edge eij∈E represents the association between features vi and vj.

[0107] After obtaining the multi-dimensional feature map, the multi-dimensional feature map can be input into the graph convolutional network for fusion, integrating data from different sources. The update rule of the graph convolutional network is:

[0108]

[0109] in, represents the activation value of the i-th neuron in the l+1th layer, σ represents the activation function, and d i and d j Respectively represent the number of connections between node i and node j, W (l) represents the weight matrix of the lth layer, represents the information of node j in the lth layer, and N(i) represents the set of neighbor nodes of node i.

[0110] After the fused data is obtained by inputting the multi-dimensional feature map into the graph convolutional network for fusion, it is also necessary to input the fused data into the convolutional neural network for feature extraction. Among them, the convolutional neural network can extract the complex patterns and structures hidden in the fused data and generate high-dimensional feature vectors. The convolutional neural network includes convolutional layers, pooling layers, and fully connected layers. After the fused data is input into the convolutional neural network, the fused data needs to be convolved first. The convolution operation formula is:

[0111] y=f(W*x+b);

[0112] Among them, * represents the convolution operation, W is the convolution kernel, x is the input feature map, b is the bias term, f() is the activation function, and the convolution result y is the extracted feature map.

[0113] Then, the feature map output by the convolution is pooled to reduce the dimension and improve the robustness of the feature. The pooling formula is:

[0114] y pool =max(x i:i+k,j:j+k );

[0115] Among them, k is the pooling window size, x i:i+k,j:j+k Represents an element within the pooling window.

[0116] In order to improve the automatic identification and screening of features related to cost prediction, give greater weight to features related to cost prediction, and improve the effectiveness and robustness of the model, the embodiment of the present application also introduces an attention mechanism, and uses the high-dimensional feature vector after attention weighting as the final output for subsequent deep learning model training. Among them, the attention weight calculation and high-dimensional feature vector calculation formula are:

[0117]

[0118] Among them, ei is the score of feature hi, hi is the i-th component of the feature vector, α i is the attention weight of feature hi. h′ is the weighted high-dimensional feature vector. The high-dimensional feature vector after attention weighting is used as the final output.

[0119] After obtaining the attention weight, the high-dimensional feature vector can be calculated. Through multi-dimensional data fusion, graph convolutional networks, convolutional neural networks and attention mechanisms, data features from different sources can be deeply fused and extracted to generate high-dimensional feature vectors with high representation capabilities.

[0120] It is understandable that high-dimensional feature vectors from different sources can also be combined to form a unified high-dimensional feature vector. For example, several high-dimensional feature vectors can be combined in a list to form a unified high-dimensional feature vector H = [h1′, h2′, h3′, h4′, …, h m ′], where hi′ is the weighted feature representation of the i-th feature source and m is the number of feature sources.

[0121] S400: Inputting the high-dimensional feature vector into the neural network model to obtain the engineering quantity interval prediction result output by the neural network model.

[0122] After obtaining the high-dimensional feature vector, the high-dimensional feature vector is input into the neural network model to obtain the engineering quantity interval prediction result. In the embodiment of the present application, the neural network model is a multi-task deep neural network model obtained by training based on the training data.

[0123] Among them, Figure 4 As shown, the steps for training a multi-task deep neural network model include:

[0124] S410: Obtain a preset amount of historical cost data.

[0125] S420: Randomly select some historical cost data as a training set.

[0126] Specifically, among the historical cost data obtained, a preset number of historical cost data are randomly selected for model training and model testing. For example, among the historical cost data, 65% of the historical cost data are randomly selected as input data for training the model, and the remaining 35% of the historical cost data can be used as a validation set for model testing.

[0127] S430: Build a multi-task deep neural network model.

[0128] The multi-task deep neural network model includes an input layer, a shared layer, a sub-network layer, a shared output layer and a sub-network output layer; the input layer is used to receive a high-dimensional feature vector, and the shared layer extracts the shared underlying features in the high-dimensional feature vector through a multi-layer fully connected network.

[0129] Among them, the input layer is mainly used to receive high-dimensional feature vectors, and the shared layer can process the input high-dimensional feature vectors through a multi-layer fully connected network to extract the shared underlying features. Each sub-network layer performs specialized predictions for a specific engineering type or region.

[0130] Specifically, Figure 5 As shown in the figure, the specific steps of building a multi-task deep neural network model include:

[0131] S4310: Set the initial network weights of the multi-task deep network model,

[0132] Specifically, the expression of the shared layer output is:

[0133] Z=σ(W s X+b s );

[0134] Among them, X is the input high-dimensional feature vector, W S is the weight matrix of the shared layer, b s is the bias vector and σ is the activation function.

[0135] The expression of the subnetwork output is:

[0136] y i =f i (W i Z+b i );

[0137] Among them, W i is the weight matrix of the sub-network, b i is the bias vector, f i is the activation function of the sub-network.

[0138] Therefore, when constructing a multi-task deep neural network model, it is necessary to first set the weights of the shared layer and the sub-network layer. Since the multi-task deep neural network model constructed in the embodiment of the present application needs to have the function of predicting engineering quantities, when setting the initial weights, it is necessary to set the initial network weights to the network weights of the pre-trained model related to the engineering quantity prediction task.

[0139] S4320: Perform fine-tuning on the initial network weights.

[0140] After setting the initial network weights, you need to fine-tune the initial network weights to adjust the weights to adapt to the new prediction task. The fine-tuned weights are:

[0141] W new =W pre +ΔW;

[0142] Among them, Wpre is the weight of the multi-task deep neural network model, and ΔW is the weight increment.

[0143] S4330: Hybrid loss function for optimizing multi-task deep neural network models.

[0144] After fine-tuning the initial network weights of the multi-task deep neural network model, it is also necessary to optimize the multi-task deep network model through a hybrid loss function to balance the prediction accuracy and robustness of each task. Different loss functions can be combined into an overall loss function to guide model training.

[0145] For example, the performance of multi-task deep neural network models is optimized by using mean squared error, L1 loss, and model regularization.

[0146] S440: Input the historical cost data into the multi-task deep neural network model, and forward propagate and back propagate according to the output results of the multi-task deep neural network model to obtain a trained multi-task deep neural network model.

[0147] Once the multi-task deep neural network model is built, the historical cost data can be input into the multi-task deep neural network model training model, and the parameters of the multi-task deep network model can be further optimized through forward propagation and back propagation of the output results, so that the multi-task deep neural network model can achieve the best performance in the predicted engineering quantity range.

[0148] Among them, forward propagation is to calculate the predicted output and loss value of the multi-task deep neural network model. Backward propagation is to calculate the gradient according to the loss value and update the model parameters through the gradient descent method.

[0149] In addition, in order to improve the prediction accuracy of the neural network model, such as Figure 6 As shown, the method for determining the engineering quantity interval of a power grid technical transformation project provided in the embodiment of the present application also includes:

[0150] S510: Adjusting the hyperparameters of the multi-task deep neural network model according to the Bayesian optimization algorithm;

[0151] Among them, hyperparameters include learning rate, number of network layers, and number of nodes; the Bayesian optimization algorithm builds a proxy model to predict the performance of different hyperparameter combinations, and gradually finds the optimal hyperparameter combination, thereby improving the performance of the multi-task deep neural network model in multi-task prediction.

[0152] Specifically, according to the Bayesian optimization algorithm, the steps of adjusting the hyperparameters of the multi-task deep neural network model include:

[0153] Construct a proxy model; wherein, in the embodiment of the present application, the proxy model uses a Gaussian process as the proxy model, wherein the prediction distribution formula of the Gaussian process is:

[0154] f(x)~GP(μ(x),k(x,x′));

[0155] Among them, f(x) represents the hyperparameter combination, μ(x) is the mean function, and k(x,x′) is the covariance function; the specific expression of the covariance function is:

[0156]

[0157] Among them, l represents the kernel width parameter.

[0158] After obtaining the proxy model, the hyperparameters of the multi-task deep neural network model can be obtained to generate preset hyperparameters according to the proxy model; and then the expected improvement value is calculated according to the preset hyperparameters; wherein the calculation formula of the expected improvement value is:

[0159] EL(x)=E[max(0),f(x + )-f(x))];

[0160] Among them, x + is the current optimal position, f(x) is the hyperparameter combination, and x is the corresponding prediction performance.

[0161] After obtaining the expected improvement value, the hyperparameter can be updated. The updated hyperparameter is x t+1 =arg maxEL(x); where t is the number of iterations.

[0162] S520: Generate engineering quantity interval results based on the engineering quantity interval prediction results according to the dynamic adjustment algorithm.

[0163] After optimizing the hyperparameters of the multi-task deep neural network model, it is also necessary to adjust the confidence interval of the engineering quantity interval prediction results.

[0164] For example, the confidence interval of the engineering quantity prediction results can be calculated based on the engineering quantity prediction results and historical cost data output by the multi-task deep neural network model. The calculation formula of the confidence interval is:

[0165]

[0166] in, is the model prediction, σ is the standard deviation of the prediction error, n is the number of samples, and z is the critical value of the standard normal distribution.

[0167] Since real-time market data, policy changes and project risk factors will affect the accuracy of the engineering quantity interval prediction results, after obtaining the confidence interval, it is also necessary to add a risk factor to the confidence interval based on the confidence interval, and flexibly expand or reduce the confidence interval according to the actual situation. Among them, adding an analysis coefficient to the confidence interval, the adjustment amount of the execution interval is:

[0168] CI adj =CI+ΔCI;

[0169] CIf inal =CI adj (1+r);

[0170] Among them, r is the risk coefficient, and ΔCI is the confidence interval adjustment calculated based on real-time factors.

[0171] On the basis of the confidence interval, the confidence interval adjustment amount is calculated according to real-time market data, policy changes and project risk factors, and then the confidence interval is adjusted. Finally, after adding the risk factor to the confidence interval, a confidence interval with a higher accuracy can be obtained.

[0172] Finally, according to the calculated confidence interval, the preset results of the engineering quantity interval are adjusted to generate the engineering quantity interval results, which can improve the accuracy and robustness of the engineering quantity interval results.

[0173] It can be seen from the above technical solutions that the present application provides a method for determining the engineering quantity interval of a power grid technical transformation project, including: obtaining the cost data of the power grid technical transformation project, the cost data including historical engineering quantity data, regional economic data, market price fluctuation data, construction environment data, and meteorological condition data; performing preprocessing on the cost data to obtain standard data; performing feature extraction on the standard data to obtain a high-dimensional feature vector; inputting the high-dimensional feature vector into a neural network model to obtain the engineering quantity interval prediction result output by the neural network model; the neural network model is a multi-task deep neural network model obtained by training based on training data, and the training data includes historical cost data. In the process of calculating the engineering quantity interval, the present application fully considers the characteristic differences in the geographical environment, climatic conditions and engineering types in different regions. Through data fusion and feature extraction, the problem of inaccurate calculation of engineering quantity intervals for power grid technical transformation projects is solved.

[0174] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the general concept of this application and do not constitute a limitation on the protection scope of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without creative work belong to the protection scope of this application.

Claims

1. A method for determining the engineering quantity interval of a power grid technical transformation project, characterized in that: include: Obtaining cost data for power grid technical transformation projects, the cost data including historical engineering quantity data, regional economic data, market price fluctuation data, construction environment data, and meteorological condition data; Preprocessing the cost data to obtain standard data; Performing feature extraction on the standard data to obtain a high-dimensional feature vector; The high-dimensional feature vector is input into a neural network model to obtain an engineering quantity interval prediction result output by the neural network model; the neural network model is a multi-task deep neural network model trained according to training data, and the training data includes historical cost data.

2. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1 is characterized in that: Preprocessing the cost data to obtain standard data includes: Performing data cleaning on the construction cost data to obtain first standard data; Normalization and linear discriminant analysis are performed on the first standard data to obtain the standard data.

3. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1 is characterized in that: Performing data cleaning on the cost data to obtain first standard data includes: Performing mean filling on the construction cost data to fill missing values ​​of the construction cost data; executing an adaptive clustering algorithm on the cost data to delete outliers in the cost data; The cost data is traversed to delete repeated values ​​in the cost data to obtain first standard data.

4. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 3 is characterized in that: Performing normalization and linear discriminant analysis on the first standard data to obtain the standard data includes: Performing minimum-maximum value normalization on the first standard data to obtain normalized data; Calculating an intra-class hash matrix and an inter-class hash matrix of the normalized data; Solving the eigenvalue and eigenvector of the product of the inverse matrix of the intra-class hash matrix and the inter-class hash matrix to obtain a projection direction; The normalized data is projected onto a projection direction to obtain standard data.

5. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1, characterized in that: Performing feature extraction on the standard data to obtain a high-dimensional feature vector includes: Constructing a multi-dimensional feature map according to the standard data; Inputting the multi-dimensional feature map into a graph convolutional network for fusion to obtain fused data; The fused data is input into a convolutional neural network for feature extraction to generate a high-dimensional feature vector; the convolutional neural network is provided with an attention mechanism.

6. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1 is characterized in that: The method further comprises: Obtaining a preset amount of the historical cost data; Randomly extracting a number of the historical cost data as a training set; Constructing a multi-task deep neural network model; the multi-task deep neural network model includes an input layer, a shared layer, a sub-network layer, a shared output layer and a sub-network output layer; the input layer is used to receive a high-dimensional feature vector, and the shared layer extracts shared underlying features in the high-dimensional feature vector through a multi-layer fully connected network; The historical cost data is input into the multi-task deep neural network model, and forward propagation and back propagation are performed according to the output results of the multi-task deep neural network model to obtain the multi-task deep neural network model that has completed training; the forward propagation is to calculate the predicted output and loss value of the multi-task deep neural network model; the back propagation is to calculate the gradient according to the loss value and update the model parameters by the gradient descent method.

7. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 6 is characterized in that: Building a multi-task deep neural network model includes: Setting the initial network weights of the multi-task deep network model, wherein the initial network weights are network weights of a pre-trained model related to the engineering quantity prediction task; performing fine-tuning on the initial network weights; The multi-task deep neural network model is optimized by a hybrid loss function.

8. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1, characterized in that: Also includes: According to the Bayesian optimization algorithm, the hyperparameters of the multi-task deep neural network model are adjusted; the hyperparameters include learning rate, number of network layers, and number of nodes; According to the dynamic adjustment algorithm, the engineering quantity interval results are generated based on the engineering quantity interval prediction results.

9. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1, characterized in that: According to the Bayesian optimization algorithm, the hyperparameters of the multi-task deep neural network model are adjusted including: Building agent models; Obtaining hyperparameters of performance data of the multi-task deep neural network to generate preset hyperparameters according to the proxy model; Calculate the expected improvement value according to the preset hyperparameters; The hyperparameter is updated according to the expected improvement value.

10. The method for determining the engineering quantity interval of a power grid technical transformation project according to claim 1, characterized in that: According to the dynamic adjustment algorithm, the engineering quantity interval results generated based on the engineering quantity interval prediction results include: Calculating the confidence interval of the preset result of the engineering quantity interval; Based on the confidence interval, a risk factor is added to adjust the preset result of the engineering quantity interval to generate an engineering quantity interval result.

Images