A method for automatically dispatching orders in power system
By adopting the automatic order dispatch method in the power system, and using TF-IDF, C4.5 decision tree, RNN and random forest algorithm to optimize the order dispatch path, the problem of inefficiency in the power system is solved, and more efficient and accurate Work order processing.
Patent Information
- Application Number
- CN202111113052.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-23
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-09-23
AI Technical Summary
There are problems in the existing power system with unreasonable collection of customer information, unscientific information transmission, repeated order assignments and insufficient service process control, resulting in low service efficiency and inability to respond quickly and solve user needs.
A power system automatic order dispatch method is adopted. By obtaining work order information, using the TF-IDF algorithm to extract keywords, build a dispatch model, and compare it with the historical dispatch information. If there is a corresponding path, the order will be automatically dispatched. If there is no Then the manual processing is transferred, and the manual results are recorded to update the model, and the single-subscribed path is optimized using C4.5 decision tree, RNN and random forest algorithm.
It improves the efficiency and accuracy of work order processing, reduces manual repeated operations, improves the automatic order dispatch rate, and provides more accurate and rapid power services.
Smart Images

Figure CN113836898B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power distribution network dispatching and operation, and in particular to an automatic dispatching method for a power system. Background Art
[0002] The popularity of smart grids and the increase in the number of users have prompted power grid companies to thrive. In this process, how to improve the quality of electricity and provide good services to electricity users has become the company's main problem. At present, there are problems in power work, such as unreasonable collection and use of customer information, unscientific transmission of customer information and needs, easy duplication of dispatch, insufficient standardized service process control, and inadequate statistical analysis of fault types.
[0003] With the acceleration of the process of urban-rural integration, the above problems will become more prominent after the establishment of the rural power repair center, which seriously restricts the work of high-quality services; according to the requirements of urban-rural integration, to improve the efficiency of power supply services, it is necessary to reduce the occurrence of power supply service difference events, and at the same time, it is necessary to prevent errors. As more and more users participate in the construction of power grid services, the current order receiving and dispatching mode can no longer meet the requirements of rapid response and rapid resolution. Although the intelligent order receiving and dispatching module of the service command platform can realize the automatic receiving and dispatching of some work orders, in the work order pool with a large demand, the number of automatic dispatches is still small. Summary of the invention
[0004] The purpose of the present invention is to provide an automatic dispatching method for an electric power system, which can respond quickly, improve the dispatching rate, and reduce manual repetitive operations.
[0005] In order to solve the above technical problems, the technical solution of the present invention is: a method for automatically dispatching orders in a power system, comprising the following steps:
[0006] Step 1: Obtain the work order call information and perform keyword extraction to obtain the work order information and build a dispatch model;
[0007] Step 2: Compare the obtained work order information with the historical dispatch information in the database. If the historical dispatch information contains a dispatch path corresponding to the current work order information, execute step 3A and then step 4. If not, transfer to manual processing and then execute step 3B, and then execute step 4.
[0008] Step 3A: dispatch orders;
[0009] Step 3B: Record and train the manual processing results, form a new dispatch path, update the dispatch model and save it in the database;
[0010] Step 4: Notify service information: After the order is dispatched, the corresponding notification information is sent to the service object.
[0011] According to the above scheme, in step 1, the work orders with incoming call information include: complaint acceptance work orders in the customer service system, dissatisfied work orders during follow-up visits, and work orders provided by the marketing system and the distribution network emergency repair platform.
[0012] According to the above scheme, the keywords include: service type, service object telephone number, account number, distribution transformer name and electricity address;
[0013] The historical dispatch information in the database includes: work order information and dispatch path; the work order information includes: service type, service object phone number, account number, distribution transformer name and electricity address; the dispatch path is the historical service log;
[0014] After receiving the order, the work order information will be compared with the work order information in the historical dispatch information. If there is a dispatch path corresponding to the same work order information, it will be automatically dispatched according to the last dispatch path. If it does not exist, it will be transferred to manual operation, and the dispatch path operated by the manual operation and the dispatch model will be updated in the database.
[0015] According to the above scheme, in step 1, the TF-IDF algorithm is used to extract keywords from the incoming call information, and the extraction steps are:
[0016] Step 1: Calculate the word frequency TF, which indicates the frequency of the entry, i.e. the keyword, appearing in the text; the word frequency TF is the ratio of the frequency of the entry to the total number of words in the file;
[0017] Step 2: Calculate the inverse document frequency (IDF). The IDF of a specific word is the logarithm of the ratio of the total number of documents to the number of documents containing the current specific word.
[0018] Step 3: Calculate TF-IDF: Multiply the term frequency TF and the inverse document frequency IDF to get TF-IDF.
[0019] According to the above scheme, in step 3A, the C4.5 decision tree algorithm is used to dispatch orders, specifically:
[0020] Determine the service type based on the service type required by the service recipient;
[0021] Arrange the distances of service outlets that can provide the current service type, and select the nearest and idle staff to serve the service recipients;
[0022] If there is incomplete information or abnormality in the outlet, it will be transferred to manual processing.
[0023] According to the above scheme, in step 3A, the Floyd algorithm is used to solve the shortest distance between the service object and the service outlet when selecting the nearest staff member.
[0024] According to the above scheme, step 3B is specifically as follows:
[0025] Step 3B.1: Extract features from the sentences of the manual processing results;
[0026] Step 3B.2: The extracted features are predicted by the constructed RNN-based classifier model to obtain the prediction results and a new distribution path; the RNN-based classifier model is trained by the LSTM algorithm;
[0027] Step 3B.3: Use the random forest algorithm to update the dispatch model and save it in the database.
[0028] According to the above scheme, the step 3B.1 is specifically as follows:
[0029] Segment the sentences of the manual processing results and use Word Embedding to represent each word as a numerical word vector;
[0030] The word vector is used to construct a sentence vector with the same dimension as the classifier model for input into the classifier model.
[0031] According to the above scheme, in step 3B.2, the classifier model structure includes:
[0032] LSTM layer: Take the sentence vector as input and use a time series LSTM operation to obtain the state vector of n hidden LSTM neural units; input X at the current moment t After entering the LSTM neural unit, the state h of the previous moment t-1 Perform calculations to obtain the state h at this moment t After inputting into the LSTM neural unit at the next moment, it is combined with the input X at the next moment t+1 Perform operations; n represents the time step, and the time step represents the number of times the recurrent neural network is expanded;
[0033] Mean pooling layer: average all state vector values;
[0034] Softmax layer: Map the state vector mean to a value of (0,1) through the softmax function to obtain the category distribution probability vector, take the category with the largest probability value as the prediction result, and obtain a new distribution path.
[0035] According to the above scheme, in step 3B.3, the method of updating the dispatch model using the random forest algorithm includes a training process and a prediction process;
[0036] The training process is:
[0037] Step S1: Given a training set S, a test set T, and a feature dimension F; determine the parameters: the number t of CART trees used, the depth d of each tree, and the number f of features used for each node; termination conditions: the minimum number s of samples on a node and the minimum information gain m on a node;
[0038] For the 1st to tth trees, i = 1 to t:
[0039] Step S2: Draw a training set S(i) with the same size as S from the training set S with replacement as the samples of the root node, and start training from the root node;
[0040] Step S3: If the termination condition is reached on the current node, set the current node as a leaf node. If it is a classification problem, the predicted output of the leaf node is the class c(j) with the largest number in the current node sample set, and the probability p is the proportion of c(j) in the current sample set; if it is a regression problem, the predicted output is the average of the sample values of the current node sample set, and then continue to train other nodes; if the termination condition is not reached on the current node, randomly select f features without replacement from the F-dimensional features; use these f features to find the one-dimensional feature k with the best classification effect and its threshold th. The samples with the kth feature less than th on the current node are divided into the left node, and the rest are divided into the right node, and continue to train other nodes;
[0041] Step S4: Repeat Step S2 and Step S3 until all nodes have been trained or marked as leaf nodes;
[0042] The prediction process is as follows:
[0043] For the 1st to tth trees, i = 1 to t:
[0044] Step Q1: Start from the root node of the current tree, and according to the threshold th of the current node, judge whether to enter the left node (<th) or the right node (>=th) until reaching a leaf node, and output the predicted value;
[0045] Step Q2: Repeat Step Q1 until the predicted values of all t trees are output; if it is a classification problem, the output is the class with the largest sum of predicted probabilities among all trees, that is, accumulate the p for each c(j); if it is a regression problem, the output is the average of the outputs of all trees.
[0046] The present invention has the following beneficial effects:
[0047] The present invention processes the received work orders automatically. When an unrecognizable work order appears, the manual operation process is recorded, i.e., the dispatch path is trained and updated to update the dispatch model. The next time the same work order appears, the dispatch path is automatically optimized to improve the efficiency of work order processing, which not only ensures the efficiency of work order processing but also improves the accuracy of work order processing. The present invention overcomes the problem of being unable to respond quickly and solve problems quickly, improves the dispatch rate of existing automatic dispatch methods, reduces manual repetitive operations, and provides more accurate and faster power services. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is an overall flow chart of an embodiment of the present invention;
[0049] Figure 2 This is a flowchart of keyword extraction in this embodiment;
[0050] Figure 3 This is a C4.5 decision tree operation flow chart in this embodiment;
[0051] Figure 4 Schematic diagram of the RNN-based classifier model in this embodiment. DETAILED DESCRIPTION
[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0053] Please refer to Figures 1 to 4 The present invention is a method for automatically dispatching orders in a power system, the steps comprising:
[0054] Step 1: Obtain the work order call information and perform keyword extraction to obtain the work order information and build a dispatch model;
[0055] In this embodiment, the TF-IDF algorithm is used to extract keywords from the acquired incoming call information. The work orders with incoming call information include: complaint acceptance work orders in the customer service system, dissatisfied work orders for return visits, and work orders provided by the marketing system and the distribution network emergency repair platform. The keywords to be extracted from these work orders include service type, service object phone number, account number, distribution transformer name, and electricity address. The extraction steps are as follows: Figure 2 As shown:
[0056] Step 1: Calculate the word frequency TF, which indicates the frequency of the entry, i.e. the keyword, appearing in the text; the word frequency TF is the ratio of the frequency of the entry to the total number of words in the file;
[0057] This number is usually normalized (usually word frequency divided by total number of words in the document) to prevent it from being biased towards long documents;
[0058]
[0059] Right now
[0060] Among them, n i,j The word is in file d j The denominator is the number of times it appears in file d j The sum of the number of occurrences of all words in ;
[0061] Step 2: Calculate the inverse document frequency (IDF). The IDF of a specific word is the logarithm of the ratio of the total number of documents to the number of documents containing the current specific word.
[0062] If there are fewer documents containing the term t and the IDF is larger, it means that the term has good category discrimination ability;
[0063]
[0064] Where |D| is the total number of documents in the corpus. |{j:t i ∈d j}| means it contains the word t i The number of files (i.e. n i,j ≠0); if the word is not in the corpus, the denominator will be zero, so 1+|{j:t i ∈d j}|, that is
[0065]
[0066] The reason why 1 is added to the denominator is to avoid the situation where the denominator is 0.
[0067] Step 3: Calculate TF-IDF: Multiply the term frequency TF and the inverse document frequency IDF to get TF-IDF;
[0068] A high word frequency within a particular document, and a low document frequency of that word in the entire document collection, can produce a high weighted TF-IDF. Therefore, TF-IDF tends to filter out common words and retain important words.
[0069] TF-IDF=TF*IDF (5)
[0070] Step 2: Compare the obtained work order information with the historical dispatch information in the database. If the historical dispatch information contains a dispatch path corresponding to the current work order information, execute step 3A and then step 4. If not, transfer to manual processing and then execute step 3B, and then execute step 4.
[0071] The database is an association database oriented to service objects. The historical order dispatch information in the database includes: order information and dispatch path; the order information includes: service type, service object's phone number, household number, distribution transformer name, and power consumption address; the dispatch path is the historical service log.
[0072] After receiving the order, compare the order information in this order with the order information in the historical order dispatch information in the association database oriented to service objects. If there is a dispatch path corresponding to the same order information, that is, if there is a historical service log for the same problem, directly perform automatic dispatch processing according to the previous dispatch path. If not, that is, no record is found, it will be transferred to the manual process, and the dispatch path operated by the manual will be updated and the dispatch model will be updated to the service object database.
[0073] Step 3A: Perform the dispatch action; when automatically performing dispatch processing with historical records, specifically use the C4.5 decision tree algorithm for dispatch, and the operation process is as Figure 3 shown. Specifically: starting from the service type required by the service object, confirm the service type; starting from the needs of the service object, arrange the distances of service outlets (power supply stations, construction teams, etc.) that can provide the current service type, and select the staff member who is the closest and in an idle state to serve the service object; if there is incomplete information or abnormal outlets, it will be transferred to the manual process for handling.
[0074] When preferentially selecting the staff member who is close in distance, use the floyd algorithm to solve the shortest distance between the service object and the service outlet. The floyd algorithm is a classic dynamic programming algorithm, which can be described in popular language as finding the shortest path from point i to point j. From the perspective of dynamic programming, a new interpretation needs to be made for this goal (this interpretation is exactly the most creative essence of dynamic programming).
[0075] The shortest path from any node i to any node j is nothing more than two possibilities. The first is directly from i to j; the second is from i passing through several nodes k to j. Therefore, the process of using the floyd algorithm to solve the shortest distance between two points is as follows:
[0076] First, assume that Dis(i,j) is the distance of the shortest path from node u to node v. Second, for each node k, check whether Dis(i,k)+Dis(k,j)<Dis(i,j) holds; if it holds, it proves that the path from i to k and then to j is shorter than the path from i directly to j, and set Dis(i,j) = Dis(i,k)+Dis(k,j). Finally, when we have traversed all nodes k, the distance recorded in Dis(i,j) is the distance of the shortest path from i to j.
[0077] Step 3B: Record and train the manual processing results, form a new dispatch path, update the dispatch model and save it in the database; when recording and training the manual processing results, use the LSTM algorithm to record and process the results, and use the random forest algorithm to update the new and old models. Use the LSTM algorithm to train the work order and save the obtained rules as a model. The model structure diagram is as follows: Figure 4 As shown, this process is actually a text classification process, which mainly consists of two parts. The first part is the sentence feature extraction of step 3B.1, and the second part is the RNN-based classifier model of steps 3B.2 and 3B.3.
[0078] Step 3B.1: Extract features from the sentences of the manual processing results;
[0079] Step 3B.1.1: Read the sentence data of the manual processing results and divide it into training set and validation set in proportion. Here, a corresponding mask vector is generated for each sentence to mark the actual length of each input text; there are several options:
[0080] (1) Considering that the words at the end of a sentence are more important, the sentence can be input in reverse order;
[0081] (2) When the number of samples is small, data augmentation can be selected, that is, the order of sentences can be disrupted to construct new samples;
[0082] (3) Sort the sentences by length
[0083] (4) Disrupt the sample order and randomly sample
[0084] Step 3B.1.2: Segment the sentences of the manual processing results and use Word Embedding to represent each word as a numerical word vector. In this process, for question texts of different lengths, pad and truncate them to the same length. If they are too short, fill them with spaces, and if they are too long, truncate them, so as to construct a model sentence vector with the same dimension as the classifier model input. The word vector is constructed into a sentence vector with the same dimension as the classifier model, which is used as the input of the classifier model;
[0085] Step 3B.2: The extracted features are predicted by the constructed RNN-based classifier model to obtain the prediction results and a new distribution path; the RNN-based classifier model is trained by the LSTM algorithm;
[0086] The classifier model structure includes LSTM layer, mean pooling layer and Softmax layer; after each word is embedded, it enters the LSTM layer. Here, the standard LSTM is used. Then, the vectors of n hidden LSTM neural units are obtained through a time series. After these vectors pass through the mean pooling layer, a state vector h can be obtained. Then, a Softmax layer is followed to obtain a category distribution probability vector. The category with the largest probability value is taken as the final prediction result. LSTM layer: With the sentence vector as input, a time series LSTM operation is performed to obtain the state vector of n hidden LSTM neural units; at the current moment, input X t After entering the LSTM neural unit, the state h of the previous moment t-1 Perform calculations to obtain the state h at this moment t After inputting into the LSTM neural unit at the next moment, it is combined with the input X at the next moment t+1 Perform operations; n represents the time step, and the time step represents the number of times the recurrent neural network is expanded;
[0087] Mean pooling layer: average all state vector values;
[0088] Softmax layer: Map the state vector mean to a value of (0,1) through the softmax function to obtain the category distribution probability vector, take the category with the largest probability value as the prediction result, and obtain a new distribution path.
[0089] Step 3B.3: Use the random forest algorithm to update the dispatch model and save it in the database.
[0090] When new data is added to the old model, that is, when manually processed dispatch information is stored in the old model, the model needs to be updated. In this embodiment, a random forest algorithm is used to implement the update between the new and old models.
[0091] Random forest refers to a classifier that uses multiple trees to train and predict samples. It is an extension of bagging. Based on the construction of bagging ensemble with decision tree as the base learner, random attribute selection is further introduced in the training process of decision tree. In simple terms, random forest is composed of multiple CART (Classification And Regression Tree). For each tree, the training set they use is sampled with replacement from the total training set, which means that some samples in the total training set may appear in the training set of a tree multiple times, or may never appear in the training set of a tree. When training the nodes of each tree, the features used are randomly extracted from all features without replacement according to a certain proportion. According to Leo Breiman's suggestion, assuming that the total number of features is M, this proportion can be sqrt(M), 1 / 2sqrt(M), 2sqrt(M).
[0092] The random forest algorithm is used to update the dispatch model, including the training process and the prediction process;
[0093] The training process is:
[0094] Step S1: Given a training set S, a test set T and a feature dimension F; determine the parameters: the number of CART trees used t, the depth of each tree d, the number of features used at each node f; termination conditions: the minimum number of samples on the node s, the minimum information gain on the node m;
[0095] For the 1-tth tree, i=1-t:
[0096] Step S2: Extract a training set S(i) with the same size as S from the training set S with replacement as the sample of the root node, and start training from the root node;
[0097] Step S3: If the termination condition is met on the current node, the current node is set as a leaf node. If it is a classification problem, the prediction output of the leaf node is the class c(j) with the largest number in the current node sample set, and the probability p is the proportion of c(j) in the current sample set; if it is a regression problem, the prediction output is the average value of each sample value in the current node sample set, and then continue to train other nodes; if the current node does not meet the termination condition, randomly select f-dimensional features from the F-dimensional features without replacement; use this f-dimensional feature to find the one-dimensional feature k with the best classification effect and its threshold th, and the samples with the k-th dimension feature of the current node that is less than th are divided into the left node, and the rest are divided into the right node, and continue to train other nodes;
[0098] Step S4: Repeat steps S2 and S3 until all nodes have been trained or marked as leaf nodes;
[0099] The prediction process is as follows:
[0100] For the 1st to t-th trees, i = 1 - t:
[0101] Step Q1: Starting from the root node of the current tree, according to the threshold th of the current node, determine whether to enter the left node (<th) or the right node (>=th) until reaching a certain leaf node, and output the prediction value;
[0102] Step Q2: Repeat Step Q1 until all t trees have output prediction values; if it is a classification problem, the output is the class with the largest sum of prediction probabilities among all trees, that is, accumulate the p for each c(j); if it is a regression problem, the output is the average of the outputs of all trees.
[0103] Step 4: Notify service information: After dispatching the order, send the corresponding notification information to the service object.
[0104] The present invention first extracts keywords from the received work order information and establishes a dispatching model; secondly, an object-oriented associated database is established. Finally, according to the historical service logs, by means of conditional screening, similar operations are automatically dispatched according to the existing processing operations, reducing manual repetitive operations. If there is no historical service log, the manual processing method is adopted, and at the same time, the manual processing process is recorded and trained to establish a new dispatching model. The present invention automatically processes the received work orders in an intelligent way. Only when the system cannot recognize a work order, this work order is transferred to manual processing. Such an approach not only ensures the work order processing efficiency but also improves the accuracy of work order processing; when the system cannot recognize a work order and is processed manually, the system can record and train the manual processing results. When there is the same work order next time, it can automatically optimize the dispatching path with the intervention of a small amount of manual work, improving the efficiency of work order processing.
[0105] Parts not involved in the present invention are the same as or implemented by the prior art.
[0106] The above content is a further detailed description of the present invention in combination with specific implementation manners. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, which should all be regarded as belonging to the protection scope of the present invention.
Claims
1. A method for automatically dispatching orders in a power system, characterized in that the steps include: Step 1: Obtain the work order call information and perform keyword extraction to obtain the work order information and build a dispatch model; The TF-IDF algorithm is used to extract keywords from the acquired call information. The work orders with call information include: complaint acceptance work orders in the customer service system, dissatisfied work orders during follow-up visits, and work orders provided by the marketing system and the distribution network emergency repair platform. The keywords that need to be extracted from these work orders include service type, service object phone number, account number, distribution transformer name, and electricity address. Step 1: Calculate the word frequency TF, which indicates the frequency of the entry, i.e. the keyword, appearing in the text; the word frequency TF is the ratio of the frequency of the entry to the total number of words in the file; Normalizing the word frequency TF to prevent the word frequency TF from being biased towards long files; Right now Among them, n i,j Is the word in file d j The denominator is the number of times it appears in file d j The sum of the number of occurrences of all words in ; Step 2: Calculate the inverse document frequency (IDF), the IDF of a word, by taking the logarithm of the ratio of the total number of documents to the number of documents containing the word; If there are fewer documents containing the term t, the IDF is larger, which means that the term has the ability to distinguish categories; Where |D| is the total number of documents in the corpus; |{j:t i ∈d j }| means containing the word t i The number of files, that is, n i,j ≠0; if the word is not in the corpus, the denominator will be zero, and 1+|{j:t i ∈d j }|, that is The reason why 1 is added to the denominator is to avoid the situation where the denominator is 0; Step 3: Calculate TF-IDF: Multiply the term frequency TF and the inverse document frequency IDF to get TF-IDF; A high word frequency within a document, and a low document frequency of the word in the entire document collection, yields a high-weighted TF-IDF, which tends to filter out common words and retain important words, i.e.: TF-IDF=TF*IDF Step 2: Compare the obtained work order information with the historical dispatch information in the database. If the historical dispatch information contains a dispatch path corresponding to the current work order information, execute step 3A and then step 4. If not, transfer to manual processing and then execute step 3B, and then execute step 4. Step 3A: dispatch orders; The C4.5 decision tree algorithm is used for dispatching orders, specifically: Determine the service type based on the service type required by the service recipient; Arrange the distances of service outlets that can provide the current service type, and select the nearest and idle staff to serve the service object; when selecting the nearest staff, use the Floyd algorithm to solve the shortest distance between the service object and the service outlet; If there is incomplete information or abnormality in the outlets, the process will be transferred to manual processing; Step 3B: Record and train the manual processing results, form a new dispatch path, update the dispatch model and save it in the database; Step 3B.1: Extract features from the sentences of the manual processing results; Step 3B.1.1: Read the sentence data of the manual processing results and divide them into training set and validation set in proportion. Here, a corresponding mask vector is generated for each sentence to mark the actual length of each input text; Step 3B.1.2: Segment the sentences of the manual processing results and use Word Embedding to represent each word as a numerical word vector. In this process, for question texts of different lengths, pad and truncate them to the same length. Short ones are padded with spaces, and long ones are truncated, so as to construct a model sentence vector with consistent dimension. The word vector is constructed into a sentence vector with the same dimension as the classifier model, which is used as the input of the classifier model; Step 3B.2: The extracted features are predicted by the constructed RNN-based classifier model to obtain the prediction results and a new distribution path; the RNN-based classifier model is trained by the LSTM algorithm; Step 3B.3: Use the random forest algorithm to update the dispatch model and save it in the database; Step 4: Notify service information: After the order is dispatched, the corresponding notification information is sent to the service object.
2. The automatic dispatching method for the power system according to claim 1, characterized in that: In step 1, the work orders with incoming call information include: complaint acceptance work orders in the customer service system, dissatisfied work orders during follow-up visits, and work orders provided by the marketing system and the distribution network emergency repair platform.
3. The automatic dispatching method for the power system according to claim 1, characterized in that: Keywords include: Service type, service recipient’s telephone number, account number, distribution transformer name and electricity usage address; The historical dispatch information in the database includes: work order information and dispatch path; the work order information includes: service type, service object phone number, account number, distribution transformer name and electricity address; the dispatch path is the historical service log; After receiving the order, the work order information will be compared with the work order information in the historical dispatch information. If there is a dispatch path corresponding to the same work order information, it will be automatically dispatched according to the last dispatch path. If it does not exist, it will be transferred to manual operation, and the dispatch path operated by the manual operation and the dispatch model will be updated in the database.
4. The automatic dispatching method for the power system according to claim 1, characterized in that: The step 3B.1 specifically includes: segmenting the sentences of the manual processing results and using Word Embedding to represent each word as a numerical word vector; The word vector is used to construct a sentence vector with the same dimension as the classifier model for input into the classifier model.
5. The automatic dispatching method for the power system according to claim 1, characterized in that: In step 3B.2, the classifier model structure includes: LSTM layer: takes the sentence vector as input and obtains the state vector of n hidden LSTM neural units through a time series LSTM operation; Mean pooling layer: average all state vector values; Softmax layer: Map the state vector mean to a value of (0,1) through the softmax function to obtain the category distribution probability vector, take the category with the largest probability value as the prediction result, and obtain a new distribution path.
6. The automatic dispatching method for the power system according to claim 1, characterized in that: In step 3B.3, the method of updating the dispatch model using the random forest algorithm includes a training process and a prediction process; The training process is: Step S1: Given a training set S, a test set T and a feature dimension F; Determine the parameters: the number of CART trees used t, the depth of each tree d, the number of features used at each node f; termination conditions: the minimum number of samples on the node s, the minimum information gain on the node m; For the 1-tth tree, i=1-t: Step S2: Extract a training set S(i) with the same size as S from the training set S with replacement as the sample of the root node, and start training from the root node; Step S3: If the termination condition is reached on the current node, the current node is set as a leaf node. If it is a classification problem, the prediction output of the leaf node is the class c(j) with the largest number in the current node sample set, and the probability p is the proportion of c(j) in the current sample set; If it is a regression problem, the prediction output is the average value of each sample value of the current node sample set, and then continue to train other nodes; if the current node does not meet the termination condition, randomly select f-dimensional features from the F-dimensional features without replacement; use these f-dimensional features to find the best one-dimensional feature k and its threshold th for classification, and the samples with the k-th dimension feature less than th on the current node are divided into the left node, and the rest are divided into the right node, and continue to train other nodes; Step S4: Repeat steps S2 and S3 until all nodes have been trained or marked as leaf nodes; The prediction process is: For the 1-tth tree, i=1-t: Step Q1: Starting from the root node of the current tree, according to the threshold th of the current node, determine whether to enter the left node or the right node until reaching a leaf node and output the predicted value; Step Q2: Repeat step Q1 until all t trees have output predicted values; if it is a classification problem, the output is the class with the largest sum of predicted probabilities among all trees, that is, the p of each c(j) is accumulated; if it is a regression problem, the output is the average of the outputs of all trees.
