Short-term power load prediction method for HPLC (High Performance Liquid Chromatography) high-frequency data acquisition

Abstract

The invention discloses a short-term power load prediction method for HPLC (High Performance Liquid Chromatography) high-frequency data acquisition. The short-term power load prediction method comprises the following steps: step a, establishing a time convolutional neural network model of a multi-attention mechanism; b, a quantile learning step based on multi-task learning; and step c, preprocessing HPLC high-frequency collected data. According to the method, deep learning and feature extraction are performed on the power load data by using the advantages of the time convolution network in the aspects of image processing and mode recognition, so that more accurate and reliable prediction is realized.

Description

Technical Field

[0001] The invention relates to a short-term power load forecasting method for HPLC high-frequency data collection in the field of power load forecasting. Background Art

[0002] Power load forecasting is a key tool for the management and operation of power systems and has become a core research area. This technology estimates future power load conditions, energy usage trends, energy demand, power consumption, etc. based on historical data records and other related variables (such as climate conditions, cultural activities, load types, time characteristics, etc.). According to the time range it covers, load forecasting is divided into short-term, medium-term, long-term and other categories. Among them, short-term forecasting can provide power companies with more accurate load information, helping them to formulate appropriate daily power generation and consumption plans, thereby enhancing their competitiveness in the power market.

[0003] At present, with the continuous advancement of the digital transformation of the power grid, its ability to collect real-time data on the use of electrical equipment is also gradually improving; for example, the use of IoT communication technology based on high-speed power line carrier communication technology (HPLC) can realize the high-frequency collection of user load data by the power consumption information collection system, which is an objective requirement and necessary communication basis for the efficient operation of the new power system. The in-depth application of HPLC can realize functions such as high-frequency data collection, active reporting of power outages, precise clock management, phase topology identification, automatic identification of substations, unified ID management, automatic synchronization of archives, communication performance monitoring and network optimization. At the same time, with the help of artificial intelligence technology and research results in related fields such as deep learning, we can deeply understand the logical relationship behind the power load statistics and better predict the possible change patterns in the future, which is also one of the very valuable research directions.

[0004] At present, the research on power grid load forecasting can be mainly divided into three directions: statistical methods based on historical data, learning methods using artificial intelligence technology (such as using neural network algorithms to process large amounts of information and make decisions), and deep strategies. The advantage of statistical methods is that they take into account the temporal correlation between data and have low computational cost, but the disadvantage is that the prediction ability for nonlinear data is limited. Load forecasting methods based on deep learning can more effectively capture the complex nonlinear characteristics of power load data and are widely used in load forecasting. Common methods include random forests, XGBoost, support vector machines, and artificial neural networks. In recent years, the emergence of emerging networks such as Transformer, Temporal Convolutional Network (TCN), and Graph Convolutional Network (GCN) has provided new ideas for multi-node load forecasting. As a temporal convolution model, TCN processes time series in parallel and can alleviate the gradient explosion problem of sequence models. As a model completely based on the self-attention mechanism, Transformer has a strong ability to extract nonlinear time series features. How to apply load forecasting methods based on deep learning to achieve accurate power load forecasting for HPLC high-frequency data acquisition is the main goal of technicians. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a short-term power load forecasting method for HPLC high-frequency data acquisition, which utilizes the advantages of temporal convolutional networks in image processing and pattern recognition to perform deep learning and feature extraction on power load data, thereby achieving more accurate and reliable predictions.

[0006] A technical solution to achieve the above purpose is: a short-term power load forecasting method for HPLC high-frequency data acquisition, comprising the following steps:

[0007] Step a, the step of establishing a temporal convolutional neural network model with a multi-head attention mechanism;

[0008] Step b: quantile learning step based on multi-task learning;

[0009] Step c, HPLC high frequency acquisition data preprocessing step.

[0010] Furthermore, in step a, the steps for establishing the temporal convolutional neural network model of the multi-head attention mechanism specifically include:

[0011] Step a01, establish the attention mechanism step;

[0012] Step a02, applying the temporal convolutional neural network step for sequence modeling tasks.

[0013] Furthermore, step a01, the steps of establishing the attention mechanism are as follows: taking a query and a set of key-value pairs as inputs, during the whole process, the query, key, value and output are represented as vectors, and the resulting output is determined by calculating the weighted sum of the values, where the weights are generated by a compatibility function that evaluates the correspondence between the query and each keyword.

[0014] Furthermore, dilated convolution is used to solve the long time series dependency problem in step a02.

[0015] Furthermore, the residual block is adopted to solve the problem of training instability and gradient vanishing caused by the increase in the number of hidden layers in the temporal convolutional neural network in step a02.

[0016] Furthermore, step b, the quantile learning step based on multi-task learning is as follows: the input data consists of two parts: historical load data (l 1 ,l 2 ,…,l t ) and historical weather data (w 1 ,w 2 ,…,w t ); first, the model adopts the attention mechanism to extract the correlation between different load data within each time granularity; then, the load and weather data of each time granularity will be connected as one of the inputs of the temporal convolutional neural network; then, the temporal convolutional neural network is used to capture the temporal correlation in the time series data as well as the uncertainty of the data and the model to achieve probabilistic load forecasting; finally, the final output ht of the temporal convolutional neural network is selected.

[0017] Furthermore, in step b, the model accelerates the training process by introducing a multi-task learning structure, extending the single quantile regression to multi-quantile regression. The output dimension of the constructed model is nx3, that is, each load needs to predict n quantiles. Regarding hyperparameter adjustment, a class called Options is defined, in which all hyperparameters are specified, and then a hyperparameter network search is performed to adjust the parameters.

[0018] Furthermore, during the model training process of step b, the gradient descent method is used to determine the optimal parameters of the model.

[0019] Further, step c, HPLC high frequency data acquisition preprocessing step specifically includes:

[0020] Step c01, data classification processing, using the model to conduct load forecasting research on representative working days and representative holiday days; when establishing the convolutional neural network load forecasting model, the first 100 weeks of data are used as training sets for convolutional neural network model training, so that it can effectively extract input data features; the last 4 weeks of data are used as test sets to verify the prediction accuracy of the model after training is completed;

[0021] Step c02, data normalization processing, linear transformation of the data through the min-max standardization method, so that the result value is converted to the interval [0, 1]. The conversion function is as follows:

[0022]

[0023] Among them, X is the original data, X min and X max are the maximum and minimum values ​​of the data respectively.

[0024] Furthermore, in step c02, for the power load, after the convolutional neural network model calculates the prediction result, it is necessary to perform denormalization to obtain the actual prediction value. The denormalization formula is as follows:

[0025] X”=(X max -X min )X'+X min .

[0026] The beneficial effects of the present invention are:

[0027] 1) In response to the growing demand for load forecasting accuracy and the interference of factors such as meteorology, economy, and date type on load forecasting, this technical solution utilizes the advantages of temporal convolutional neural network (TCN) in dealing with nonlinear problems. By analyzing the characteristics of load data, a TCN-based power load forecasting model is established to improve forecasting accuracy.

[0028] 2) In order to solve the problems of large fluctuations in the prediction output curve of the traditional BP neural network model and obvious deviations between the prediction results and the actual values, this technical solution establishes a TCN-based power load prediction model and uses a time convolutional neural network (TCN) to perform load prediction. The results are compared with the prediction results of the BP neural network model. The results show that TCN for HPLC high-frequency data collection can well predict the load data for the next 24 hours, and is more accurate than the BP neural network prediction results.

[0029] The present invention uses a time convolutional neural network to establish short-term load forecasting and ultra-short-term load forecasting models, and uses the advantages of TCN in image processing and pattern recognition to perform deep learning and feature extraction on power load data to achieve more accurate and reliable predictions. Comparative analysis of the BP neural network model prediction in dealing with nonlinear problems, output curve fluctuation amplitude, average relative error, model stability and other performance shows that TCN is more accurate and stable in the field of short-term and ultra-short-term load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of the multiplication (dot product) attention function;

[0031] Figure 2 This is a diagram of the configuration of the multi-head attention mechanism;

[0032] Figure 3 This is the TCN network structure diagram;

[0033] Figure 4 This is the network structure diagram of the expanded causal convolution;

[0034] Figure 5 Schematic diagram of residual block structure;

[0035] Figure 6 Schematic diagram of pinball loss function;

[0036] Figure 7 This is a diagram of a quantile regression model based on multi-task learning;

[0037] Figure 8 This is a graph showing the power load forecast results for a typical working day in the embodiment;

[0038] Fig. 9 This is a graph showing the power load forecast results for a typical rest day in the embodiment;

[0039] Fig.10 This is a graph showing the load forecasting results of the BP neural network during working days in the embodiment;

[0040] Fig.11 It is a comparison chart of short-term prediction and ultra-short-term prediction of the embodiment. DETAILED DESCRIPTION

[0041] In order to better understand the technical solution of the present invention, the following is a detailed description through specific embodiments:

[0042] The present invention discloses a short-term power load forecasting method based on time convolutional neural network for HPLC high-frequency data collection. The method is based on time convolutional neural network (TCN) and specifically includes:

[0043] Step a, the step of establishing a temporal convolutional neural network model with a multi-head attention mechanism;

[0044] Step b: quantile learning step based on multi-task learning;

[0045] Step c, HPLC high frequency acquisition data preprocessing step.

[0046] The steps of establishing the TCN model of the multi-head attention mechanism in step a are specifically as follows:

[0047] Step a01, establish the attention mechanism step;

[0048] The attention mechanism is a neural network architecture that can capture the correlation between elements even without a clear topological structure. Since there is no clear topological relationship between various loads in the same area, but there is some correlation between them, it can be used to extract the correlation between different loads.

[0049] The attention mechanism takes a query and a set of key-value pairs as input. Throughout the process, the query, key, value, and output are represented as vectors. The resulting output is determined by calculating the weighted sum of the values. These weights are generated by a compatibility function that evaluates the correspondence between the query and each keyword. The two most commonly used attention mechanisms are additive attention and multiplicative (dot product) attention. In the present invention, multiplicative (dot product) attention is used, which excels in computational speed and space efficiency in real-world applications. The schematic diagram of the multiplicative (dot product) attention function is shown in Figure 1. Figure 1 shown.

[0050] The present invention adopts a multi-head attention architecture instead of using a single attention function with key, value and query. Compared with the single attention mechanism, this approach allows the model to simultaneously focus on multiple reproduction subspaces at different positions. The configuration of the multi-head attention mechanism is as follows Figure 2 shown.

[0051] Step a02, applying the temporal convolutional neural network step for sequence modeling tasks.

[0052] The temporal convolutional neural network (TCN) combines the characteristics of recurrent neural network (RNN) and CNN. It is a dilated causal convolution that can be used for sequence modeling tasks. The output length of the TCN network is the same as the input length.

[0053]

[0054] Where: x i is the input of the model network; is the predicted value of the model network.

[0055] In order to solve the problem of long time series dependency, TCN uses dilated convolution. The idea of ​​dilated convolution is to add holes to traditional convolution and skip part of the input by sampling intervals to increase the size of the receptive field, thereby using a smaller convolution kernel size and a smaller number of layers to achieve modeling of long time series dependency. The network structure of TCN is as follows Figure 3 shown.

[0056] x t is the input feature of the model at time t, y tis the predicted value of the model at time t. The figure takes the convolution kernel size of 3 and the network depth of 3 as an example to draw the calculation method of the last output node. d is the dilation factor size, which indicates the step size between calculation nodes. The result of the next layer in the model is the input of the previous layer after the dilation convolution operation. For a one-dimensional vector input x∈Rn and a convolution filter f∶{0,1,...,k-1}→R, the calculation formula of the dilation convolution is as follows

[0057]

[0058] Where: d is the dilation factor; *d is the convolution symbol indicating dilated convolution; x* d f is the dilated convolution with the dilation factor d of x and f; k is the filter size of the convolution kernel, and the dilation factor is equivalent to using a fixed step size between two adjacent parts of the convolution kernel. When d=1, the dilated convolution degenerates into a normal convolution. Residual connections are also used in the TCN network structure, which directly adds the input x to the output result of the next layer, which can effectively avoid the gradient explosion and gradient disappearance problems in deep neural networks.

[0059] x l+1 =x l +F(x l )

[0060] Where: x l is the input of the lth layer; x l+1 is the output of layer l+1; F is the nonlinear transformation in the neural network, such as convolution operation.

[0061] Standard causal convolutions have limitations when examining past information, mainly due to the fixed receptive field of the convolution kernel. In tasks that require important historical data, networks consisting only of standard causal convolutions have greater difficulties, including increasing the network depth, which may lead to slow training. To effectively solve the problem of processing large amounts of historical data, TCN introduces an extended causal convolution structure that expands the convolution kernel to effectively expand the receptive field, namely dilated convolution. Extended causal convolution allows the network to capture patterns on various time scales and capture long-range dependencies. For a one-dimensional time series input X = (x 0 ,x 1 ,...,x t ) and convolution kernel f:{0,2,...,x n-1}, the dilated convolution operation on the sequence element t is as follows:

[0062]

[0063] Where n is the kernel size, d is the dilation factor, and x t-d·i Indicates the direction of the past.

[0064] In order to enable TCN to process input sequences of different lengths and convert them into output sequences of consistent length, TCN introduces padding, which is defined as follows.

[0065] p=(n-1)·d

[0066] Where p is the padding size, n is the kernel size, and d is the dilation factor. In this way, TCN has the same ability to process time series data as LSTM.

[0067] The structure of the dilated causal convolution is as follows Figure 4 As shown, where the kernel size n=2 and the dilation factor d=[1, 2, 4].

[0068] As the number of hidden layers in TCN increases, the kernel size and dilation factor increase, and the receptive field of TCN becomes larger, capturing more information. However, the increase in the number of layers in TCN will lead to problems such as training instability and gradient vanishing. To solve this problem, the present invention adopts a residual block, whose structure is as follows Figure 5 shown.

[0069] A residual block consists of two branches. One branch performs the transformation operation f(·), which contains a dilated causal convolution layer, a weight normalization layer, a ReLU activation function, and a dropout layer. The dilated causal convolution layer is constructed by combining the causal convolution and the dilated convolution mentioned above. These layers are used to extract hidden features from the input. The weight normalization layer speeds up training by constraining the weight range. The activation function is performed using a ReLU selected for efficient convergence. To alleviate the overfitting problem in deep networks, the dropout layer is used as a regularization technique.

[0070] The other branch is a 1X 1 convolutional layer, which ensures the number of feature maps next to the first branch is consistent. (h-1) The output of the h-th residual block X is as follows:

[0071] X (h) =σ(X (h-1) +f(X (h-1) ))

[0072] where f(·) is an activation function.

[0073] More specifically, if the entire TCN consists of M residual blocks, f(·) is ignored and the forward propagation of the entire TCN is expressed as follows.

[0074]

[0075] Where X (0) is the input of TCN, X (M) is the output of TCN. Thus, we calculate TCN to X (0)The formula for the total loss of back propagation is as follows.

[0076]

[0077] Where L is the total loss of the entire TCN, and "1" means that the gradient at the TCN output can be directly propagated back to the TCN input, which reduces the probability of the gradient vanishing problem. In fact, it can be used for any pair (x i ,x j ), where x i and x j They are the i-th and j-th residual blocks of TCN, respectively, thus solving the gradient disappearance problem.

[0078] Step b: The quantile learning steps based on multi-task learning are as follows:

[0079] b01)Quantile regression;

[0080] There are many forms of probabilistic prediction, and the present invention adopts quantile regression as a prediction method. On the one hand, quantile regression is a non-parametric method that allows direct probabilistic prediction without assuming the distribution of the prediction target. On the other hand, by utilizing quantiles, the complete probability distribution function of the prediction target can be derived. Quantile regression approximates the conditional distribution of a random variable by estimating the quantiles and is expressed as an optimization task aimed at minimizing the pinball loss function. The pinball loss is a weighted absolute error for any quantile q∈(0,1) as follows:

[0081]

[0082] Where y is the target value and y is the quantile prediction at q. The pinball loss function penalizes the model's prediction deviation at a specific q. When y is greater than f, the penalty term is q; conversely, when f is greater than y, the penalty term is q-1. More specifically, the pinball loss function is as follows Figure 6 shown.

[0083] b02) Multi-task learning;

[0084] The disadvantage of quantile regression is that a specific model needs to be built and trained for each quantile of the predicted distribution to be published. 1 ,q 2 ,...,q s}, to predict the entire quantile set, it is necessary to train s models, which easily leads to a large number of models being used to construct the entire prediction distribution, reducing computational efficiency. In this regard, the present invention adopts a multi-task learning structure. Multi-task learning allows the prediction model to use multiple pinball loss functions at the same time, thereby achieving simultaneous prediction of multiple quantiles. In addition, in multi-task learning, various subtasks use standard feature sharing layers with the same feature parameters. In this way, during the back propagation process of the neural network, these shared layers can be optimized simultaneously, thereby reducing computational costs.

[0085] b03) Quantile regression based on multi-task learning;

[0086] The proposed model is Figure 7 As shown, the input data consists of two parts: historical load data (l 1 ,l 2 ,…,l t ) and historical weather data (w 1 ,w 2 ,…,w t ). First, the model uses an attention mechanism to extract the correlation between different load data within each time granularity. Then, the load and weather data of each time granularity will be connected as one of the inputs of TCN. TCN is then used to capture the temporal correlation in time series data and the uncertainty of data and model to achieve probabilistic load forecasting. Finally, the final output ht of TCN is selected. The model extends single quantile regression to multi-quantile regression by introducing a multi-task learning structure to accelerate the training process. The output dimension of the constructed model is nx3, that is, each load needs to predict n quantiles. Regarding hyperparameter adjustment, a class called Options is defined, in which all hyperparameters are specified, and then a hyperparameter network search is performed to adjust the parameters. During the model training process, the gradient descent method is used to determine the optimal parameters of the model.

[0087] The step c, HPLC high frequency data acquisition preprocessing step is specifically:

[0088] Step c01, data classification processing:

[0089] The embodiment uses detailed power load data from 2020 to 2021 in the region, and the total data set covers a time span of 104 weeks. In order to obtain more detailed load change information, high-frequency sampling is performed at 1 hour, and finally a huge data set of up to 17,472 samples is obtained.

[0090] In order to comprehensively evaluate the prediction performance of the TCN model under different circumstances, we specially selected Monday as a representative of working days and used the model to study the load forecasting on working days. At the same time, we also selected Sunday as a representative of holidays to explore the performance of the model in holiday load forecasting.

[0091] When establishing the TCN load forecasting model, the first 100 weeks of data were used as the training set for TCN model training, so that it can effectively extract the input data features. The last 4 weeks of data were used as the test set to verify the prediction accuracy of the model after training.

[0092] Step c02, data normalization processing:

[0093] The data is linearly transformed using the min-max standardization method so that the resulting value is converted to the interval [0, 1]. The conversion function is as follows:

[0094]

[0095] Among them, X is the original data, X min and X max are the maximum and minimum values ​​of the data respectively.

[0096] For power load, after the TCN model calculates the prediction result, it needs to be denormalized to obtain the actual prediction value. The denormalization formula is as follows:

[0097] X”=(X max -X min )X'+X min

[0098] The following simulation prediction is carried out by combining the actual load data of a certain area to verify the effectiveness of the model.

[0099] Given that the time span of short-term power load forecasting is relatively short, there is no need to set too many convolutional layers and pooling layers when constructing the TCN model to avoid overfitting. A TCN load forecasting model with two layers of convolutional layers and pooling layers arranged alternately is constructed. To ensure that the model is fully trained, the amount of data used in each iteration is set to 18, and the maximum number of training rounds is 30. During the training process, the real-time training progress display function is used to keep track of the model training status at any time. After the training is completed, the initial prediction results are denormalized to obtain the final applicable power load forecast value.

[0100] The analysis steps of this embodiment include:

[0101] a) TCN working day load forecast analysis;

[0102] b) TCN rest day load forecast analysis;

[0103] c) Comparative analysis between TCN and BP neural network;

[0104] d) Ultra-short-term load forecasting analysis based on TCN

[0105] The TCN short-term load forecasting analysis for HPLC high-frequency load data is specifically as follows:

[0106] a) Select the load data of a certain area every Monday from 2020 to 2021, take the first 100 days as the training samples of the input convolutional neural network, and predict the power load on Monday of the next week. The results are as follows Figure 8 shown.

[0107] Figure 8 The actual load values ​​and TCN model prediction values ​​for each hour on Monday are given and the relative errors are calculated. Analysis shows that the maximum relative error of the TCN model prediction is 2.75%, the minimum is 0.03%, and the average error is 1.14%. The overall error is small, indicating the effectiveness of TCN in predicting short-term loads.

[0108] b) Select the load data of a region every Sunday from 2020 to 2021, take the first 100 days as the input TCN training sample, and predict the power load of the next Sunday. The results are as follows Fig. 9 shown.

[0109] Fig. 9 The actual load value and TCN model prediction value for each hour on Sunday are given, and the relative error is calculated. It can be seen that in the TCN model prediction results, the maximum relative error is 4.13%, the minimum is 0.05%, the average error is 1.19%, and the overall error is small. According to the curve in the figure, in the first 15 hours, the TCN model prediction results are very close to the true value, almost consistent, while the prediction deviation is large around 18:00, mainly due to the large uncertainty of the evening peak load, which affects the overall prediction accuracy to a certain extent.

[0110] Comprehensive predictions for weekdays and weekends show that the difference between the maximum and minimum relative errors of weekday predictions is smaller, while the difference between the maximum and minimum relative errors of weekend predictions is larger, and the average relative errors of the two are approximately equal.

[0111] c) To verify the superiority of the TCN model, this technical solution substitutes the workday load data of a certain area into the BP neural network model for prediction and compares the results with Fig.10 Compare the results in .

[0112] Fig.10 Given the actual load values ​​for each hour on Monday and the predicted values ​​of the BP neural network model, it can be seen that the maximum relative error of the prediction results is 4.7%, the minimum is 0.07%, and the average relative error is as high as 1.68%, which is higher than the TCN model, showing a significant gap. Through the above in-depth analysis, it can be seen that the performance of the TCN model in short-term load forecasting is superior to that of the BP neural network model.

[0113] d) In order to expand the application of TCN in the field of power load forecasting, ultra-short-term load forecasting is performed. First, the 24-hour load data is extracted at intervals of 15 minutes, and there are 96 load data in a day. Then, the data of the first 100 Mondays from 2020 to 2021 in a certain area are used as training sets to predict the load data of the next Monday.

[0114] Fig.11 The results of the ultra-short-term load forecasting using the TCN model are given. It can be seen that the prediction results of the TCN model are very close to the actual values ​​before 8:00, and there are large errors after 13:00, mainly due to the large uncertainty of the load in the latter part. By comparison, it is found that the average relative error of the short-term prediction error of TCN for most regions at ultra-short-term prediction errors is 0.84%, which is 0.29% lower than the short-term prediction, demonstrating the applicability of TCN in dealing with ultra-short-term load forecasting problems.

[0115] The present invention builds a TCN load forecasting model, and substitutes the HPLC high-frequency collected data for simulation experiments, and conducts load forecasting for working days and rest days respectively. The results show that TCN can better complete the short-term load forecasting task, with an average relative error of only 1.13% and 1.17%. It fully proves the effectiveness of TCN in short-term load forecasting. In order to further verify the superiority of the TCN model, its working day forecasting results are also compared with the BP neural network model forecasting results. It is found that the TCN model has obvious advantages in both forecasting accuracy and stability, providing support for expanding the application of TCN in the field of power load forecasting.

[0116] Those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A short-term power load forecasting method for HPLC high-frequency data acquisition, characterized in that: The steps include: Step a, the step of establishing a temporal convolutional neural network model with a multi-head attention mechanism; Step b: quantile learning step based on multi-task learning; Step c, HPLC high frequency acquisition data preprocessing step.

2. A short-term power load forecasting method for HPLC high-frequency data collection according to claim 1, characterized in that: Step a: The steps for establishing the temporal convolutional neural network model with a multi-head attention mechanism include: Step a01, establish the attention mechanism step; Step a02, applying the temporal convolutional neural network step for sequence modeling tasks.

3. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 2 is characterized in that: Step a01, the specific steps of establishing the attention mechanism are: taking the query and a set of key-value pairs as input. Throughout the process, the query, key, value and output are represented as vectors. The resulting output is determined by calculating the weighted sum of the values. These weights are generated by the compatibility function that evaluates the correspondence between the query and each keyword.

4. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 2 is characterized in that: Use dilated convolution to solve the long time series dependency problem in step a02.

5. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 4 is characterized in that: The residual block is used to solve the problem of training instability and gradient disappearance caused by the increase in the number of hidden layers in the temporal convolutional neural network in step a02.

6. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 1 is characterized in that: Step b, the quantile learning step based on multi-task learning is as follows: the input data consists of two parts: historical load data (l1, l2, ..., l t ) and historical weather data (w1,w2,…,w t ); first, the model adopts the attention mechanism to extract the correlation between different load data within each time granularity; then, the load and weather data of each time granularity will be connected as one of the inputs of the temporal convolutional neural network; then, the temporal convolutional neural network is used to capture the temporal correlation in the time series data as well as the uncertainty of the data and the model to achieve probabilistic load forecasting; finally, the final output ht of the temporal convolutional neural network is selected.

7. A short-term power load forecasting method for HPLC high-frequency data collection according to claim 6, characterized in that: In step b, the model accelerates the training process by introducing a multi-task learning structure, extending the single quantile regression to multi-quantile regression. The output dimension of the constructed model is nx3, that is, each load needs to predict n quantiles. Regarding hyperparameter adjustment, a class called Options is defined, in which all hyperparameters are specified, and then a hyperparameter network search is performed to adjust the parameters.

8. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 7 is characterized in that: During the model training process in step b, the gradient descent method is used to determine the optimal parameters of the model.

9. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 1 is characterized in that: Step c, HPLC high frequency data acquisition preprocessing step specifically includes: Step c01, data classification processing, using the model to conduct load forecasting research on representative working days and representative holiday days; when establishing the convolutional neural network load forecasting model, the first 100 weeks of data are used as training sets for convolutional neural network model training, so that it can effectively extract input data features; the last 4 weeks of data are used as test sets to verify the prediction accuracy of the model after training is completed; Step c02, data normalization processing, linear transformation of the data through the min-max standardization method, so that the result value is converted to the interval [0, 1]. The conversion function is as follows: Among them, X is the original data, X min and X max are the maximum and minimum values ​​of the data respectively.

10. The short-term power load forecasting method for HPLC high-frequency data collection according to claim 1, characterized in that: In step c02, for the power load, after the convolutional neural network model calculates the prediction result, it is necessary to perform denormalization to obtain the actual prediction value. The denormalization formula is as follows: X”=(X max -X min )X'+X min 。

Images