A method for dynamic line capacity expansion based on multi-source data fusion

Abstract

The present invention discloses a method for dynamic capacity increase of a line based on multi-source data fusion. The invention has the effect of establishing a neural network embedding layer in a neural network structure model based on a probabilistic sparse self-attention mechanism by fusing multi-source information such as feature sequence information and hierarchical time series information, and using these variables for modeling to extract the periodic law of capacity change of transmission lines, and to achieve long-term series prediction based on historical sequences, so as to reasonably predict the trend of future working condition changes. The advantage of the present invention is that it can realize the prediction and judgment within a period of time in the future. Through the judgment within a period of time, the comprehensive changes of a certain line or multiple lines within a period of time in the future can be known, which is conducive to the better and more reasonable allocation of the rated load of the transmission line in advance.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems and relates to a method for dynamically increasing the capacity of high-voltage lines, and in particular to a method for deep learning modeling of dynamic capacity increase of lines based on multi-source data fusion. Background Art

[0002] At present, the dynamic capacity expansion technology of power transmission lines is relatively mature. It mainly monitors the power transmission lines and environmental conditions. Under the premise of not breaking the current technical regulations, the maximum allowable current carrying capacity of the conductor is calculated according to the current carrying capacity model such as Morgan, so as to make full use of the objective implicit capacity of the line and improve the actual transmission capacity of the transmission line. The real-time monitoring parameters include the temperature, tension, sag, temperature, sunshine, customs and other parameter characteristics of the conductor. According to the real-time above parameter characteristics, the maximum allowable current carrying capacity of each transmission line is changed in time, so as to give full play to the performance of each line, so as to transmit a larger load of electricity and improve the compensation capacity of the transmission capacity of the transmission line.

[0003] At present, with the rise of new energy and the advancement of science and technology, more new energy power plants, including solar energy and wind energy, have been built in the western region of my country. The transmission characteristics of these new energy sources are that they belong to renewable energy, but they also bring some problems. Due to the long distance and high energy, although additional transmission lines can be added, they must also be carried by existing transmission lines. So whether such high-energy new energy electricity can be borne by the load capacity of existing transmission lines, and at the same time, it is also necessary to consider how to more efficiently utilize and allocate the transmission capacity of existing transmission lines.

[0004] In the process of dynamic capacity expansion of transmission lines, the existing technology mostly adopts single working condition prediction, such as the patent with publication number CN110321601A, which discloses a method and system for advanced prediction of dynamic current carrying capacity of overhead lines. It predicts by introducing a recurrent neural network with attention mechanism. The result is that it can only predict one or several points in the future. For example, it predicts a line point in the next 15 minutes and the capacity required by this line at that time node. This prediction method cannot perform iterative learning. In the era of multi-source power data fusion, such prediction alone obviously cannot achieve the purpose of prediction. For example, it is necessary to predict the rated load on multiple lines in the next hour so as to make reasonable deployment during high voltage and peak hours of large-scale power transmission. Summary of the invention

[0005] The purpose of the present invention is to enable multiple transmission lines to simultaneously predict the load capacity within a period of time in the future, so as to achieve the technical problem of dynamic capacity expansion effect under the state of multi-source power data fusion. A line dynamic capacity expansion method and system based on multi-source data fusion is designed, and the capacity of the transmission line is modeled by a neural network structure based on a probabilistic sparse self-attention mechanism. The training process is divided into two processes: pre-training and hyperparameter fine-tuning. First, based on the historical operation data of multiple lines of the same voltage level in the regional power grid, the neural network model is pre-trained to extract the fuzzy correlation between each input feature variable, between the input feature and the prediction target, and between the prediction targets. Then, the pre-trained model is used as the initial parameter, and the model network hyperparameters are fine-tuned based on the historical operation data of the target line. In the forward prediction stage of the model, long-term series prediction is realized based on the historical sequence, and the future trend of working condition changes is reasonably predicted.

[0006] The technical solution of the present invention is a method for dynamic capacity increase of a line based on multi-source data fusion, which predicts the capacity of a future transmission line by building a prediction model. The key is that the specific steps of the method for dynamic capacity increase of a line are as follows:

[0007] a. Construct a neural network structure model with a probabilistic sparse self-attention mechanism, add a model embedding layer consisting of hierarchical time series information, feature sequence information, and sequence position information, and use the model embedding layer as the input sample set of the pre-training model to form a pre-training model;

[0008] b. Extract historical operation data of multiple lines of the same voltage level, and use the pre-trained model for pre-training to form a correlation between the model embedding layer and the transmission line capacity;

[0009] c. When the brigade's target transmission line is predicted, the target transmission line is trained using the pre-trained model according to the hierarchical time series information of the target transmission line and the relevant information of the target variable, and the final target line dynamic capacity increase prediction model is obtained;

[0010] d. Use the trained line dynamic capacity increase prediction model to make long-sequence predictions on the capacity of transmission lines at multiple times in the future.

[0011] In the step a, the hierarchical time series information includes season, month, week, day, and time information.

[0012] In step a, the characteristic sequence information includes temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, and historical line capacity.

[0013] In the step b, the association between the model embedding layer and the transmission line capacity includes the association between each input feature variable and between the input feature and the prediction target.

[0014] In the step c, the hierarchical time series information of the target transmission line and the relevant information operation data of the target variable include season, month, week, day, time information as well as temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, and historical line capacity.

[0015] The beneficial effects of the present invention are:

[0016] 1. Based on the historical operation data of multiple lines of the same voltage level in the regional power grid, the neural network model is pre-trained. When one of the lines needs to be evaluated, the parameters of this line can be input, and the pre-trained model can optimize this line. Because there are records of pre-training, it can be continuously iterated and trained. Through the model, deep learning can be continuously carried out to achieve near-perfect predictive processing capabilities, ultimately making the prediction results more accurate.

[0017] 2. The present invention can realize the prediction and judgment of a period of time in the future. Through the judgment of a period of time, the comprehensive changes of a certain line or multiple lines in the future can be known, which is conducive to the better and more reasonable allocation of the rated load of the transmission line in advance.

[0018] 3. In the neural network structure model based on the probabilistic sparse self-attention mechanism, a neural network embedding layer is established by integrating multi-source information such as feature sequence information and hierarchical time series information. These variables are used for modeling to extract the periodic law of transmission line capacity changes, realize long-term series prediction based on historical sequences, and make reasonable predictions on future operating condition changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is an embedded layer structure of the line dynamic capacity increase model of multi-source data fusion in the present invention. DETAILED DESCRIPTION

[0020] The detailed steps of the method for dynamically increasing the capacity of a line according to the present invention are described below in conjunction with the accompanying drawings and specific embodiments.

[0021] The focus of the present invention is on the selection of models and the improvement of models. The basic model adopts a neural network structure based on a probabilistic sparse self-attention mechanism, and establishes a neural network embedding layer to pre-train the model by fusing multi-source information such as feature sequence information, hierarchical time series information, and sequence position information. The addition of the embedding layer makes the prediction model analysis more comprehensive, and can comprehensively analyze the relationship between variable factors affecting line capacity. The hierarchical time series information includes: season, month, week, day, and time information, and the feature sequence information includes: temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, and historical line capacity. It is used to extract the periodic law of transmission line capacity changes.

[0022] In the model pre-training stage, the neural network model is pre-trained based on the historical operation data of multiple lines of the same voltage level in the power grids of various regions to extract the fuzzy correlation between the input feature variables, between the input features and the prediction targets, and between the prediction targets.

[0023] In the target variable training phase of a certain line, the pre-trained model is used as the initial parameter, and the model is fine-tuned according to the hierarchical timing information of the target transmission line and the target variable related information operation data;

[0024] In the forward prediction stage of the model, long-term series prediction is achieved based on historical sequences to make reasonable predictions on future trends in operating conditions.

[0025] The basic model is a neural network structure based on the probabilistic sparse self-attention mechanism. The input data includes: season, month, week, day, time information, temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, historical line capacity; the output data is: transmission line capacity. The specific steps of the probabilistic sparse self-attention mechanism of the basic model are as follows:

[0026] Step 1. Determine the input: Q∈R m×d , K∈R n×d , V∈R n×d , where m and n are the input sequence lengths, and d is the model input dimension;

[0027] Step 2: Set the impact factor c, calculate the sampling dimension u=c·lnm of Q, and the randomly selected sampling dimension U=c·lnn of K;

[0028] Step 3: Randomly select U keys from K to form Calculate the dot product score

[0029] Step 4: Calculate measurement indicators by row Based on M, extract the top-u queries from Q to form

[0030] Step 5: Calculate the dot product of u queries and all keys, that is

[0031] Step 6: Calculate S0=mean(V), where S={S1,S0} is the output of the probabilistic sparse self-attention.

[0032] In order to comprehensively analyze the relationship between variable factors affecting line capacity, a model embedding layer construction method of multi-source data fusion of feature sequence information, feature sequence position information, and hierarchical temporal information is proposed. The model embedding layer includes hierarchical temporal information, sequence information, and sequence position information, such as Figure 1 .

[0033] By observing the curves of line capacity and its characteristic influencing factors, it is found that its changing pattern has a strong time periodicity. The hierarchical time series information such as season, month, week, day, and hour is extracted, encoded and added to the model embedding layer, and the hierarchical time series information extracted through Informer model training is used to mine the impact of time period on line capacity. The characteristic sequence information consists of temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, and historical line capacity, corresponding to u0-u6 in the figure. The sequence position information E0-E6 represents the relative position of these variables in the encoding process, which helps to deepen the intrinsic connection between characteristic variables and line capacity. The multi-source data fusion expression of characteristic sequence information, characteristic sequence position information, and hierarchical time series information is as follows:

[0034]

[0035] in: Represents the variables of the multi-source data fusion embedding layer; represents the sequence characteristic variable; pe represents the position information of the sequence characteristic variable; se represents the hierarchical time series information, pos represents the sequence position, i represents the model dimension position, t represents the sample working condition, L x Represents the input sample length. The feature variable position encoding sequence formula is as follows:

[0036] pe(pos,2i)=sin(pos / 10000 2i / d )

[0037] pe(pos,2i+1)=cos(pos / 10000 2i / d )

[0038] Where pos represents the sequence position and i represents the model dimension position.

[0039] During the model pre-training process, transmission lines with certain similarities in operating conditions are selected and trained based on multi-source information.

[0040] In the process of model fine-tuning, based on the results of the pre-trained model, the model is further trained on the historical data of the target line to make the model more adaptable to the operating conditions of the target line.

[0041] The long-term series prediction process can achieve the prediction of the future long-term change trend of line capacity by setting hyperparameters such as the length of the historical sequence and the prediction target prediction length. The model decomposes the input of the decoding layer into two parts, namely, the partial historical sequence and the target placeholder sequence, to achieve one-step prediction of the long sequence. The formula is as follows:

[0042]

[0043] Among them, the historical sequence L token is the length of the history sequence; the target placeholder sequence L y The sequence data of the predicted target sequence length is set to 0. The sparse self-attention in the decoding layer adopts a mask mechanism, and the dot product of the mask is set to -inf, which can prevent each position from paying attention to future positions, thereby avoiding autoregression.

Claims

1. A method for dynamic capacity increase of power lines based on multi-source data fusion, which predicts the capacity of future power transmission lines by building a prediction model, and is characterized by: The specific steps of the line dynamic capacity increasing method are as follows: a. Construct a neural network structure model with a probabilistic sparse self-attention mechanism, add a model embedding layer consisting of hierarchical time series information, feature sequence information, and sequence position information, and use the model embedding layer as the input sample set of the pre-training model to form a pre-training model; b. Extract historical operation data of multiple lines of the same voltage level, and use the pre-trained model for pre-training to form a correlation between the model embedding layer and the transmission line capacity; c. When the brigade's target transmission line is predicted, the target transmission line is trained using the pre-trained model according to the hierarchical time series information of the target transmission line and the relevant information of the target variable, and the final target line dynamic capacity increase prediction model is obtained; d. Use the trained line dynamic capacity increase prediction model to make long-sequence predictions on the capacity of transmission lines at multiple times in the future. In the step a, the model embedding layer includes hierarchical time series information, sequence information and sequence position information. The hierarchical time series information of season, month, week, day and hour is extracted and encoded and added to the model embedding layer. The hierarchical time series information extracted by Informer model training mines the influence of time period on line capacity. The sequence information consists of temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity and historical line capacity. The multi-source data fusion expression of sequence information, sequence position information and hierarchical time series information is as follows: in: Represents the variables of the multi-source data fusion embedding layer; represents the sequence characteristic variable; pe represents the position information of the sequence characteristic variable; se represents the hierarchical time series information, pos represents the sequence position, i represents the model dimension position, t represents the sample working condition, L x Represents the input sample length, where the feature variable position encoding sequence formula is as follows: pe(pos,2i)=sin(pos / 100002i / d) pe(pos,2i+1)=cos(pos / 100002i / d) Where pos represents the sequence position, i represents the model dimension position, and d is the model input dimension.

2. The method for dynamic line capacity expansion based on multi-source data fusion according to claim 1, characterized in that: In the step b, the association between the model embedding layer and the transmission line capacity includes the association between each input feature variable and between the input feature and the prediction target.

3. The method for dynamic line capacity expansion based on multi-source data fusion according to claim 1, characterized in that: In the step c, the hierarchical time series information of the target transmission line and the relevant information operation data of the target variable include season, month, week, day, time information as well as temperature, ambient humidity, wind speed, wind direction, rainfall, radiation intensity, and historical line capacity.

Images