A method and system for predicting power material demand based on deep learning

Abstract

The present invention relates to the technical field of material demand forecasting, and in particular to a method and system for forecasting electric power material demand based on deep learning, including using convolutional neural networks to extract features from electric power engineering drawings to obtain unstructured feature representations of electric power materials; using natural language processing technology to construct a multidimensional feature vector of electric power engineering material demand based on the unstructured feature representations of electric power materials and annotation information of electric power engineering drawings; using a spectral clustering algorithm to classify electric power material demand based on the similarity measurement of the multidimensional feature vectors to form sets of electric power material demand of different categories; using a material demand forecasting model to forecast material demand for each set of electric power material demand, and obtaining material demand forecasting results. The present invention realizes efficient feature extraction and multidimensional feature fusion of electric power engineering drawings through deep learning technology, thereby improving the accuracy and efficiency of electric power material demand forecasting and enhancing the intelligent level of material management.

Description

Technical Field

[0001] The present invention relates to the technical field of material demand forecasting, and in particular to a method and system for forecasting power material demand based on deep learning. Background Art

[0002] With the continuous development of the power industry, the demand forecast of power materials has received more and more attention, and the coordinated relationship between the supply and demand of power materials directly affects the development of the entire power industry. However, the existing power material demand forecasting methods have encountered many challenges in practical applications, which seriously limit the improvement of their prediction accuracy and effective support for decision-making.

[0003] Existing technologies mainly rely on historical data and structured information to forecast material demand. Although this method can reflect the historical laws of material demand to a certain extent, it ignores other major factors that may affect material demand, especially those contained in unstructured data. For example, engineering drawings are an important basis for power design and construction, which contain a large amount of key information such as equipment configuration and material selection. This information plays a vital role in accurately predicting material demand. However, existing models have not been able to effectively mine and utilize these unstructured data, resulting in deviations in prediction results. Due to the lack of processing capabilities for unstructured data, existing models cannot fully capture the complexity and diversity of material demand, thereby affecting the accuracy of predictions; secondly, when processing material demand, existing technologies fail to fully consider the actual demand differences between different materials. There are many types of materials in the power industry, and each material often has significant differences in demand and demand laws. However, existing models often use a unified prediction method and parameter setting, and fail to perform classified predictions based on the characteristics of different materials, thereby affecting the accuracy and pertinence of the prediction results.

[0004] In summary, these problems existing in the existing power material demand forecasting methods limit the effectiveness of existing models in practical applications. Therefore, a more comprehensive and accurate power material demand forecasting method is urgently needed to improve the accuracy and practicality of power material demand forecasting. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method and system for predicting power material demand based on deep learning.

[0006] In a first aspect, the present invention provides a method for predicting power material demand based on deep learning, the method comprising the following steps:

[0007] Obtaining electric power engineering drawings, extracting features from the electric power engineering drawings using a convolutional neural network, and obtaining unstructured feature representations of electric power materials;

[0008] Parsing the annotation information of the power engineering drawings, and constructing a multi-dimensional feature vector of power engineering material demand using natural language processing technology based on the unstructured feature representation of the power materials and the annotation information;

[0009] According to the similarity measurement of multi-dimensional feature vectors, the spectral clustering algorithm is used to classify the power material demand to form a set of power material demand of different categories;

[0010] The pre-built material demand forecasting model is used to forecast the material demand of each category of power material demand sets to obtain the material demand forecasting results.

[0011] In a further embodiment, the step of extracting features from the electric power engineering drawings using a convolutional neural network to obtain unstructured feature representations of electric power materials includes:

[0012] Converting the power engineering drawings of different formats into the format specification of the power engineering drawings to obtain pre-processed power engineering drawings of a unified format;

[0013] Converting the pre-processed electric power engineering drawing into an image format to obtain an electric power engineering drawing image;

[0014] A convolutional neural network is used to extract features from the electric power engineering drawing image to obtain an unstructured feature representation of electric power materials.

[0015] In a further embodiment, the convolutional neural network adopts a ResNet-50 convolutional neural network.

[0016] In a further embodiment, the step of constructing a multidimensional feature vector of power engineering material demand using natural language processing technology based on the unstructured feature representation of power materials and the annotation information includes:

[0017] Performing named entity recognition on the unstructured feature representation of the power materials to extract unstructured key information of material demand;

[0018] Perform text mining on the annotation information of the power engineering drawings, calculate the TF-IDF value of each word in the annotation information, and filter out the material demand annotation keywords according to the TF-IDF value;

[0019] Conduct correlation analysis between the unstructured key information of material demand and the key words of material demand annotation, and use the co-occurrence matrix to calculate the confidence between the two;

[0020] According to the confidence level, a target unstructured feature set satisfying a confidence threshold is selected from the unstructured feature representation of the electric power material;

[0021] Convert the target unstructured feature set into structured data to obtain the target structured feature set;

[0022] The template matching method is used to extract the structural feature representation of power materials from the power engineering drawing image;

[0023] The target structured feature set and the structured feature representation of power materials are uniquely encoded to form a multi-dimensional feature vector of power engineering material demand.

[0024] In a further implementation scheme, the step of performing named entity recognition on the unstructured feature representation of the power materials to extract unstructured key information of material demand includes:

[0025] Using a support vector machine to perform binary classification on the unstructured feature representation of the power materials, and generate a preliminary material requirement table for the power engineering drawings;

[0026] Named entity recognition is performed on the preliminary material demand table to extract unstructured key information of the material demand.

[0027] In a further implementation scheme, the unstructured key information of material demand is obtained by performing named entity recognition on the preliminary material demand table using a bidirectional long short-term memory network-conditional random field model.

[0028] In a further embodiment, the step of classifying the power material demand by using a spectral clustering algorithm according to the similarity measurement of the multidimensional feature vector to form a set of power material demand of different categories includes:

[0029] Performing standardization processing on the multidimensional feature vector to obtain a set of standardized multidimensional feature vectors;

[0030] Calculating the similarity between every two different feature vectors in the standardized multidimensional feature vector set by using cosine similarity, and screening out feature vector pairs whose similarity is greater than a preset similarity threshold;

[0031] Taking each feature vector in the feature vector pair as a node, and constructing an edge between each pair of the feature vectors according to the similarity, to obtain a weighted similarity network;

[0032] The spectral clustering algorithm is used to perform cluster analysis on the weighted similarity network to form different categories of power material demand sets.

[0033] In a further embodiment, the process of constructing the material demand forecasting model includes:

[0034] Extracting historical power material demand data of various types of power materials within a preset time range, and dividing the historical power material demand data into time granularity to obtain historical power material demand patterns under different time periods;

[0035] The historical demand pattern of power materials in each time period is divided into spatial granularity to obtain the historical spatial demand distribution of power materials at different spatial granularities;

[0036] Extracting spatial demand characteristics from the historical spatial demand distribution of power materials using a spatial autocorrelation analysis method;

[0037] Combining the spatial demand characteristics according to time granularity to form a historical spatiotemporal demand change sequence of power materials at different spatiotemporal granularities;

[0038] The empirical mode decomposition is used to decompose the historical spatiotemporal demand change sequence of each category of power materials at multiple scales, and the demand fluctuation sequence at different spatiotemporal scales is obtained.

[0039] The historical spatiotemporal demand change sequence of electric power materials and the corresponding demand fluctuation sequence are used as input features, and a material demand forecasting model for different types of electric power materials is obtained by training a convolutional neural network.

[0040] In a further embodiment, the total loss function of the material demand forecasting model is specifically:

[0041]

[0042] in,

[0043]

[0044]

[0045]

[0046] In the formula, is the total loss function of the material demand forecasting model; is the regression error of the power material demand set of the i-th category; M is the total number of categories of the power material demand set; is the weight coefficient of the spatiotemporal series prediction error; is the prediction error of the spatiotemporal series; is the weight coefficient of the forecast error of the demand fluctuation series; is the forecast error of the demand fluctuation sequence; J is the demand quantity in the i-th category of power material demand set; is the actual power material demand in the power material demand set; The predicted power material demand in the power material demand set; is the number of spatial regions; is the number of time periods; is the real space-time demand value of the nth spatial region in the tth time period; is the predicted spatiotemporal demand value of the nth spatial region in the tth time period; is the number of demand fluctuation series; is the true value of the kth fluctuation sequence of the mth type of power materials; It is the predicted value of the kth fluctuation sequence of the mth type of electric power materials.

[0047] In a second aspect, the present invention provides a power material demand forecasting system based on deep learning, the system comprising:

[0048] A feature extraction module is used to obtain electric power engineering drawings, extract features from the electric power engineering drawings using a convolutional neural network, and obtain unstructured feature representations of electric power materials;

[0049] A feature analysis module, used to parse the annotation information of the power engineering drawings, and construct a multi-dimensional feature vector of the power engineering material demand using natural language processing technology based on the unstructured feature representation of the power materials and the annotation information;

[0050] The classification module is used to classify the power material demand according to the similarity measurement of the multi-dimensional feature vector and use the spectral clustering algorithm to form a set of power material demand of different categories;

[0051] The demand forecasting module is used to use the pre-built material demand forecasting model to forecast the material demand of each category of power material demand sets and obtain the material demand forecasting results.

[0052] The present invention provides a method and system for forecasting power material demand based on deep learning. The method obtains power engineering drawings, extracts features from the power engineering drawings using convolutional neural networks, and obtains unstructured feature representations of power materials; parses annotation information of the power engineering drawings, and constructs a multidimensional feature vector of power engineering material demand based on the unstructured feature representations of power materials and the annotation information using natural language processing technology; classifies power material demand based on the similarity measurement of the multidimensional feature vector using a spectral clustering algorithm to form power material demand sets of different categories; and predicts material demand for each category of power material demand set using a pre-constructed material demand prediction model to obtain material demand prediction results. Compared with the prior art, the method realizes efficient feature extraction and multidimensional feature fusion of power engineering drawings through deep learning technology, thereby improving the accuracy and efficiency of power material demand forecasting and the intelligent level of material management. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of a flow chart of a method for predicting power material demand based on deep learning provided by an embodiment of the present invention;

[0054] Figure 2 This is a block diagram of a power material demand forecasting system based on deep learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0056] refer to Figure 1 , the embodiment of the present invention provides a method for predicting power material demand based on deep learning, such as Figure 1 As shown, the method comprises the following steps:

[0057] S1. Obtain electric power engineering drawings, use convolutional neural networks to extract features from the electric power engineering drawings, and obtain unstructured feature representations of electric power materials.

[0058] In this embodiment, the step of extracting features from the electric power engineering drawings using a convolutional neural network to obtain unstructured feature representations of electric power materials includes:

[0059] Converting the power engineering drawings of different formats into the format specification of the power engineering drawings to obtain pre-processed power engineering drawings of a unified format;

[0060] Converting the pre-processed electric power engineering drawing into an image format to obtain an electric power engineering drawing image;

[0061] A convolutional neural network is used to extract features of the electric power engineering drawing image to obtain an unstructured feature representation of electric power materials; wherein the convolutional neural network adopts a ResNet-50 convolutional neural network.

[0062] Specifically, this embodiment can receive the power engineering drawing files uploaded or passed in by the user through the user interface or automated process. These power engineering drawing files can be files of different formats such as AutoCAD files, PDF files, JPEG / PNG image files, etc. In order to facilitate processing, this embodiment determines a unified drawing format according to the standards of the power industry, and uses a format conversion tool or library to convert the collected power engineering drawings into a unified format. For example, for power engineering drawings identified as CAD format, the system selects a CAD parsing library (such as Teigha, Open CASCADE, etc.) for parsing. Using the selected CAD parsing library, the system converts the CAD binary file into a parsable DXF (Drawing Exchange Format, Drawing Exchange Format) file, DXF is an open text format, which is convenient for subsequent information extraction and processing. At the same time, in order to use convolutional neural networks for feature extraction, the system needs to convert the pre-processed power engineering drawings after format conversion into image formats (such as PNG, JPEG), and then select ResNet-50 as the convolutional neural network for feature extraction according to task requirements. ResNet-50 is a deep residual network with powerful feature extraction capabilities. In this embodiment, the pre-trained weights of ResNet-50 can be loaded from the pre-trained model library, and the last layer of ResNet-50 is removed, and it is retained as a feature extractor. In this embodiment, the pre-processed power engineering drawing image is input into the ResNet-50 model, and feature vectors are obtained by forward propagation calculation. These feature vectors are the unstructured feature representations of power materials. In this embodiment, power engineering drawings of different formats are received and converted into image formats suitable for convolutional neural network processing, and then the unstructured feature representations of power materials are extracted to provide a data basis for subsequent analysis.

[0063] S2. parse the annotation information of the power engineering drawings, and construct a multi-dimensional feature vector of the power engineering material demand using natural language processing technology based on the unstructured feature representation of the power materials and the annotation information.

[0064] In this embodiment, the step of constructing a multidimensional feature vector of power engineering material demand using natural language processing technology according to the unstructured feature representation of power materials and the annotation information includes:

[0065] Performing named entity recognition on the unstructured feature representation of the power materials to extract unstructured key information of material demand;

[0066] Perform text mining on the annotation information of the power engineering drawings, calculate the TF-IDF value of each word in the annotation information, and filter out the material demand annotation keywords according to the TF-IDF value;

[0067] Conduct correlation analysis between the unstructured key information of material demand and the key words of material demand annotation, and use the co-occurrence matrix to calculate the confidence between the two;

[0068] According to the confidence level, a target unstructured feature set satisfying a confidence threshold is selected from the unstructured feature representation of the electric power material;

[0069] Convert the target unstructured feature set into structured data to obtain the target structured feature set;

[0070] The template matching method is used to extract the structural feature representation of power materials from the power engineering drawing image;

[0071] The target structured feature set and the structured feature representation of power materials are uniquely encoded to form a multi-dimensional feature vector of power engineering material demand.

[0072] The step of performing named entity recognition on the unstructured feature representation of the electric power materials to extract unstructured key information of material demand includes:

[0073] Using a support vector machine to perform binary classification on the unstructured feature representation of the power materials, and generate a preliminary material requirement table for the power engineering drawings;

[0074] Named entity recognition is performed on the preliminary material demand table to extract unstructured key information of the material demand.

[0075] Specifically, this embodiment uses a support vector machine (SVM) algorithm to perform binary classification on the unstructured feature representation of power materials to generate a preliminary material demand table, and then uses the named entity recognition (NER) technology in natural language processing to parse the text content in the preliminary material demand table through a BiLSTM-CRF model (bidirectional long short-term memory network-conditional random field model) to identify key information such as equipment name, model, quantity, and the identified named entities are extracted as unstructured key information of material demand. At the same time, this embodiment preprocesses the annotation information of the power engineering drawings, including word segmentation, stop word removal, etc., and uses the TF-IDF (term frequency-inverse document frequency) algorithm to calculate the weight of each word to evaluate its importance in the annotation information. The size of the F-IDF value is used to filter out the keywords most relevant to the material demand, and form the material demand annotation keywords. Then, this embodiment counts the co-occurrence of the unstructured key information of the material demand and the material demand annotation keywords in the drawings, constructs a co-occurrence matrix, and uses the data in the co-occurrence matrix to calculate the confidence between each unstructured key information of the material demand and the material demand annotation keyword to evaluate the degree of association between them, and then filters out the features whose correlation degree (confidence) with the material demand annotation keywords is higher than the preset confidence threshold from the unstructured feature representation of the power materials to form a target unstructured feature set, converts the features in the target unstructured feature set into a structured data format, such as a key-value pair, a table, etc., and organizes the converted data to form a target structured feature set.

[0076] At the same time, this embodiment sets a template for matching the power material information in the power engineering drawings according to the characteristics and rules of the power engineering drawings, so as to adopt a sliding window method, search for the area matching the template in the drawing image through the template matching method, and extract the structured feature representation of the power materials. Finally, by one-hot encoding (One-Hot Encoding) each feature in the target structured feature set and the structured feature representation of the power materials, it is converted into a binary vector, and the one-hot encoded feature vectors are concatenated to form a multidimensional feature vector representing the power engineering material demand, thereby realizing the conversion of unstructured features and annotation information in the power engineering drawings into multidimensional feature vectors, providing support for subsequent analysis and prediction.

[0077] S3. Based on the similarity measurement of the multidimensional feature vector, the spectral clustering algorithm is used to classify the power material demand to form a set of power material demand of different categories.

[0078] In this embodiment, the step of classifying the power material demand by using the spectral clustering algorithm according to the similarity measurement of the multidimensional feature vector to form different categories of power material demand sets includes:

[0079] Performing standardization processing on the multidimensional feature vector to obtain a set of standardized multidimensional feature vectors;

[0080] Calculating the similarity between every two different feature vectors in the standardized multidimensional feature vector set by using cosine similarity, and screening out feature vector pairs whose similarity is greater than a preset similarity threshold;

[0081] Taking each feature vector in the feature vector pair as a node, and constructing an edge between each pair of the feature vectors according to the similarity, to obtain a weighted similarity network;

[0082] The spectral clustering algorithm is used to perform cluster analysis on the weighted similarity network to form different categories of power material demand sets.

[0083] Specifically, after obtaining the multidimensional feature vector of the power engineering material demand, since each dimensional feature vector in the multidimensional feature vector has different dimensions and value ranges, it is necessary to perform standardization processing on it. Specifically, the mean and standard deviation of each dimensional feature vector are calculated, and then the value of each dimensional feature vector is subtracted from the mean and divided by the standard deviation to obtain a standardized multidimensional feature vector set. Then, the cosine similarity formula is used to calculate the similarity between every two different feature vectors in the standardized multidimensional feature vector set. The cosine similarity is an indicator to measure the degree of similarity between the directions of two vectors, and its value is between -1 and 1. When the directions of the two vectors are exactly the same, the cosine similarity is 1; when the directions are completely opposite, the cosine similarity is -1; when the two vectors are orthogonal (i.e., vertical), the cosine similarity is 0. This embodiment sets a similarity threshold according to actual needs. This similarity threshold is used to screen out those feature vector pairs with higher similarity, so as to subsequently construct a weighted similarity network. The method is to compare the calculated similarity with a preset similarity threshold, screen out all feature vector pairs whose similarity is greater than the similarity threshold, and use each feature vector in the screened feature vector pairs as a node of a weighted similarity network. According to the similarity between the feature vector pairs, an edge is constructed between each pair of feature vectors. The weight of the edge is the similarity value between the feature vector pairs. All nodes and edges are combined to form a weighted similarity network. This weighted similarity network reflects the similarity relationship between the feature vectors. Finally, this embodiment uses the weighted similarity network as the input of the spectral clustering algorithm, and performs clustering analysis through the spectral clustering algorithm. The spectral clustering algorithm divides the data points into different clusters according to the similarity relationship between the feature vectors, and converts the clustering results into different categories of power material demand sets. The power material demand in each category has similar feature vectors and similar demand patterns. These power material demand sets can better understand the distribution and characteristics of power material demand.

[0084] S4. Use the pre-built material demand forecasting model to forecast the material demand of each category of power material demand sets to obtain the material demand forecasting results.

[0085] In this embodiment, the process of constructing the material demand forecasting model includes:

[0086] Extracting historical power material demand data of various types of power materials within a preset time range, and dividing the historical power material demand data into time granularity to obtain historical power material demand patterns under different time periods;

[0087] The historical demand pattern of power materials in each time period is divided into spatial granularity to obtain the historical spatial demand distribution of power materials at different spatial granularities;

[0088] Extracting spatial demand characteristics from the historical spatial demand distribution of power materials using a spatial autocorrelation analysis method;

[0089] Combining the spatial demand characteristics according to the time granularity to form a historical spatiotemporal demand change sequence of power materials at different spatiotemporal granularities;

[0090] The empirical mode decomposition is used to decompose the historical spatiotemporal demand change sequence of each category of power materials at multiple scales, and the demand fluctuation sequence at different spatiotemporal scales is obtained.

[0091] The historical spatiotemporal demand change sequence of electric power materials and the corresponding demand fluctuation sequence are used as input features, and a material demand forecasting model for different types of electric power materials is obtained by training a convolutional neural network.

[0092] Specifically, this embodiment can extract historical power material demand data of various types of power materials (such as transformers, cables, switches, etc.) within a preset time range in the past from the power database, and divide the historical power material demand data according to different time periods (such as year, quarter, month, week, day) to obtain demand data under different time periods, and analyze to obtain demand patterns under different time scales, and then divide the historical power material demand patterns into spaces according to the geographical areas of the power grid, divide the demand data by spatial areas in each time period, and analyze to obtain demand distribution under different spatial granularities, and then this embodiment adopts a spatial autocorrelation analysis method (such as Moran's I index) calculates the demand correlation between different spatial regions, extracts spatial demand features with significant spatial correlation according to the results of spatial autocorrelation analysis, combines the extracted spatial demand features according to the time granularity, forms a power material demand change sequence at different spatiotemporal granularities, performs empirical mode decomposition (EMD) on each category of power material spatiotemporal demand change sequence, decomposes it into multiple intrinsic mode functions (IMFs) and a residual, each IMF represents a demand fluctuation sequence at different spatiotemporal scales, takes the power material spatiotemporal demand change sequence and the corresponding demand fluctuation sequence as input features, uses a convolutional neural network (CNN) to train the input features, and obtains a material demand forecasting model for different power material categories. In this embodiment, the model performance can be evaluated by cross-validation and other methods, and the parameters and structure of the CNN model can be adjusted according to the evaluation results to improve the prediction accuracy. In this embodiment, the material demand forecasting model of this embodiment comprehensively considers the prediction errors of the spatiotemporal sequence and the fluctuation sequence, thereby constructing a more comprehensive and targeted loss function, thereby improving the performance of the model in the power material demand forecasting scenario. The total loss function of the material demand forecasting model is specifically:

[0093]

[0094] in,

[0095]

[0096]

[0097]

[0098] In the formula, is the total loss function of the material demand forecasting model; is the regression error of the power material demand set of the i-th category; M is the total number of categories of the power material demand set; is the weight coefficient of the spatiotemporal series prediction error; is the prediction error of the spatiotemporal series; is the weight coefficient of the forecast error of the demand fluctuation series; is the forecast error of the demand fluctuation sequence; J is the demand quantity in the i-th category of power material demand set; is the actual power material demand in the power material demand set; The predicted power material demand in the power material demand set; is the number of spatial regions; is the number of time periods; is the real space-time demand value of the nth spatial region in the tth time period; is the predicted spatiotemporal demand value of the nth spatial region in the tth time period; is the number of demand fluctuation series; is the true value of the kth fluctuation sequence of the mth type of power materials; It is the predicted value of the kth fluctuation sequence of the mth type of electric power materials.

[0099] The embodiment of the present invention provides a method for forecasting power material demand based on deep learning, the method comprising obtaining power engineering drawings, extracting features from the power engineering drawings using a convolutional neural network, and obtaining unstructured feature representations of power materials; parsing the annotation information of the power engineering drawings, and constructing a multidimensional feature vector of power engineering material demand based on the unstructured feature representations of power materials and the annotation information using natural language processing technology; classifying power material demand based on the similarity measurement of the multidimensional feature vector using a spectral clustering algorithm to form power material demand sets of different categories; and forecasting the material demand of each category of power material demand sets using a pre-constructed material demand forecasting model to obtain a material demand forecasting result. Compared with the prior art, the method realizes efficient feature extraction and multidimensional feature fusion of power engineering drawings through deep learning technology, thereby improving the accuracy and efficiency of power material demand forecasting, improving the intelligent level of material management, and being able to optimize resource allocation and reduce costs.

[0100] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0101] In one embodiment, Figure 2 As shown, an embodiment of the present invention provides a power material demand forecasting system based on deep learning, and the system includes:

[0102] The feature extraction module 101 is used to obtain electric power engineering drawings, extract features from the electric power engineering drawings using a convolutional neural network, and obtain unstructured feature representations of electric power materials;

[0103] The feature analysis module 102 is used to parse the annotation information of the power engineering drawings, and construct a multi-dimensional feature vector of the power engineering material demand by using natural language processing technology according to the unstructured feature representation of the power materials and the annotation information;

[0104] The classification module 103 is used to classify the power material demand according to the similarity measurement of the multi-dimensional feature vector and to form a set of power material demand of different categories by using the spectral clustering algorithm;

[0105] The demand forecasting module 104 is used to use a pre-built material demand forecasting model to perform material demand forecasting on each category of power material demand sets to obtain a material demand forecasting result.

[0106] For the specific limitations of a power material demand forecasting system based on deep learning, please refer to the above-mentioned limitations of a power material demand forecasting method based on deep learning, which will not be repeated here. A person of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0107] The embodiment of the present invention provides a power material demand forecasting system based on deep learning. The system obtains power engineering drawings through a feature extraction module, extracts features from the power engineering drawings using a convolutional neural network, and obtains unstructured feature representations of power materials; the feature analysis module parses the annotation information of the power engineering drawings, and constructs a multidimensional feature vector of power engineering material demand based on the unstructured feature representation of power materials and the annotation information using natural language processing technology; the classification module classifies the power material demand based on the similarity measurement of the multidimensional feature vector and uses a spectral clustering algorithm to form different categories of power material demand sets; the demand forecasting module uses a pre-built material demand forecasting model to forecast material demand for each category of power material demand sets, and obtains material demand forecasting results. Compared with the prior art, the system realizes efficient feature extraction and multidimensional feature fusion of power engineering drawings through deep learning technology, thereby improving the accuracy and efficiency of power material demand forecasting, improving the intelligent level of material management, and optimizing resource allocation and reducing costs.

[0108] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A method for predicting power material demand based on deep learning, characterized in that: The following steps are involved: Obtaining electric power engineering drawings, extracting features from the electric power engineering drawings using a convolutional neural network, and obtaining unstructured feature representations of electric power materials; Parsing the annotation information of the power engineering drawings, and constructing a multi-dimensional feature vector of power engineering material demand using natural language processing technology based on the unstructured feature representation of the power materials and the annotation information; According to the similarity measurement of multi-dimensional feature vectors, the spectral clustering algorithm is used to classify the power material demand to form a set of power material demand of different categories; The pre-built material demand forecasting model is used to forecast the material demand of each category of power material demand sets to obtain the material demand forecasting results.

2. A method for predicting power material demand based on deep learning as claimed in claim 1, characterized in that: The step of extracting features from the electric power engineering drawings using a convolutional neural network to obtain unstructured feature representations of electric power materials includes: Converting the power engineering drawings of different formats into the format specification of the power engineering drawings to obtain pre-processed power engineering drawings of a unified format; Converting the pre-processed electric power engineering drawing into an image format to obtain an electric power engineering drawing image; A convolutional neural network is used to extract features from the electric power engineering drawing image to obtain an unstructured feature representation of electric power materials.

3. The method for predicting power material demand based on deep learning according to claim 2, characterized in that: The convolutional neural network adopts the ResNet-50 convolutional neural network.

4. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The step of constructing a multidimensional feature vector of power engineering material demand using natural language processing technology according to the unstructured feature representation of power materials and the annotation information comprises: Performing named entity recognition on the unstructured feature representation of the power materials to extract unstructured key information of material demand; Perform text mining on the annotation information of the power engineering drawings, calculate the TF-IDF value of each word in the annotation information, and filter out the material demand annotation keywords according to the TF-IDF value; Conduct correlation analysis between the unstructured key information of material demand and the key words of material demand annotation, and use the co-occurrence matrix to calculate the confidence between the two; According to the confidence level, a target unstructured feature set satisfying a confidence threshold is selected from the unstructured feature representation of the electric power material; Convert the target unstructured feature set into structured data to obtain the target structured feature set; The template matching method is used to extract the structural feature representation of power materials from the power engineering drawing image; The target structured feature set and the structured feature representation of power materials are uniquely encoded to form a multi-dimensional feature vector of power engineering material demand.

5. A method for predicting power material demand based on deep learning as claimed in claim 4, characterized in that: The step of performing named entity recognition on the unstructured feature representation of the electric power materials to extract unstructured key information of material demand includes: Using a support vector machine to perform binary classification on the unstructured feature representation of the power materials, and generate a preliminary material requirement table for the power engineering drawings; Named entity recognition is performed on the preliminary material demand table to extract unstructured key information of the material demand.

6. The method for predicting power material demand based on deep learning according to claim 5, characterized in that: The unstructured key information of material demand is obtained by performing named entity recognition on the preliminary material demand table using a bidirectional long short-term memory network-conditional random field model.

7. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The step of classifying the power material demand by using the spectral clustering algorithm according to the similarity measurement of the multidimensional feature vector to form different types of power material demand sets includes: Performing standardization processing on the multidimensional feature vector to obtain a set of standardized multidimensional feature vectors; Calculating the similarity between every two different feature vectors in the standardized multidimensional feature vector set by using cosine similarity, and screening out feature vector pairs whose similarity is greater than a preset similarity threshold; Taking each feature vector in the feature vector pair as a node, and constructing an edge between each pair of the feature vectors according to the similarity, to obtain a weighted similarity network; The spectral clustering algorithm is used to perform cluster analysis on the weighted similarity network to form different categories of power material demand sets.

8. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The construction process of the material demand forecasting model includes: Extracting historical power material demand data of various types of power materials within a preset time range, and dividing the historical power material demand data into time granularity to obtain historical power material demand patterns under different time periods; The historical demand pattern of power materials in each time period is divided into spatial granularity to obtain the historical spatial demand distribution of power materials at different spatial granularities; Extracting spatial demand characteristics from the historical spatial demand distribution of power materials using a spatial autocorrelation analysis method; Combining the spatial demand characteristics according to the time granularity to form a historical spatiotemporal demand change sequence of power materials at different spatiotemporal granularities; The empirical mode decomposition is used to decompose the historical spatiotemporal demand change sequence of each category of power materials at multiple scales, and the demand fluctuation sequence at different spatiotemporal scales is obtained. The historical spatiotemporal demand change sequence of electric power materials and the corresponding demand fluctuation sequence are used as input features, and a material demand forecasting model for different types of electric power materials is obtained by training a convolutional neural network.

9. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The total loss function of the material demand forecasting model is specifically: in, In the formula, is the total loss function of the material demand forecasting model; is the regression error of the power material demand set of the i-th category; M is the total number of categories of the power material demand set; is the weight coefficient of the spatiotemporal series prediction error; is the prediction error of the spatiotemporal series; is the weight coefficient of the forecast error of the demand fluctuation series; is the forecast error of the demand fluctuation sequence; J is the demand quantity in the i-th category of power material demand set; is the actual power material demand in the power material demand set; The predicted power material demand in the power material demand set; is the number of spatial regions; is the number of time periods; is the real space-time demand value of the nth spatial region in the tth time period; is the predicted spatiotemporal demand value of the nth spatial region in the tth time period; is the number of demand fluctuation series; is the true value of the kth fluctuation sequence of the mth type of power materials; It is the predicted value of the kth fluctuation sequence of the mth type of electric power materials.

10. A power material demand forecasting system based on deep learning, characterized in that: The system comprises: A feature extraction module is used to obtain electric power engineering drawings, extract features from the electric power engineering drawings using a convolutional neural network, and obtain unstructured feature representations of electric power materials; A feature analysis module, used to parse the annotation information of the power engineering drawings, and construct a multi-dimensional feature vector of the power engineering material demand using natural language processing technology based on the unstructured feature representation of the power materials and the annotation information; The classification module is used to classify the power material demand according to the similarity measurement of the multi-dimensional feature vector and use the spectral clustering algorithm to form a set of power material demand of different categories; The demand forecasting module is used to use the pre-built material demand forecasting model to forecast the material demand of each category of power material demand sets and obtain the material demand forecasting results.

Images