Power load prediction method based on deep cross-temporal convolution network (DTCN) model

Abstract

The application provides a power load prediction method based on a deep cross-time convolution network (DTCN) model, which is used for predicting ultra-short-term power load. The method uses continuous and discrete feature vectors in a data set, extracts data features through automatic feature engineering by using DCN, inputs the arranged dense vectors in time sequence to a TCN for prediction, respectively uses sigmoid and tanh activation functions and expands causal convolution to fit parameters, establishes a model, and finally decodes to obtain an ultra-short-term prediction result of the load. The application has low data capacity requirement, fast prediction speed and high precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load prediction, in particular to a power load prediction method based on a deep cross-time convolution network (DTCN) model. BACKGROUND

[0002] With the reform of the spot market, the behaviors of signing, pricing, and economic calculation based on load prediction make the research on load prediction a hot topic in energy research. Power load data is a necessary component to ensure the balance between power generation and demand. Power grid operators need high-precision load prediction to maintain the safety and stability of power supply. Accurate and effective short-term load prediction can achieve energy saving, cost reduction, and fine scheduling management of smart grids, and early warning to ensure safety. Accurate prediction of residential electricity demand can balance the supply and demand relationship during peak electricity consumption, reducing the waste of power resources and having a positive impact on the protection of the ecological environment. Artificial neural network algorithms construct network frameworks layer by layer for high-precision prediction, have high fault tolerance for sample data, strong non-linear mapping capability, self-adaptation, and self-organization, and are widely used in the field of power load prediction, becoming one of the research hotspots. Common short-term load prediction methods include traditional time series prediction models, traditional machine learning methods, and deep learning methods that have emerged in recent years. These prediction methods have achieved certain results, but still have challenges. The main limitations are:

[0003] 1) These networks have strong specificity, small application scope, and weak generalization ability;

[0004] 2) The processing of discrete features (category features) is usually simple, mainly through direct input and one-hot processing;

[0005] 3) The modeling process is complex and difficult to efficiently handle variable influencing factors. SUMMARY

[0006] The present application proposes a power load prediction method based on a deep cross-time convolution network (DTCN) model, which has low data capacity requirements, fast prediction speed, and high precision.

[0007] The present application adopts the following technical solutions.

[0008] The power load prediction method based on the deep cross-time convolution network (DTCN) model is used for ultra-short-term power load prediction. The method uses continuous and discrete feature vectors in the data set, extracts data features through DCN automatic feature engineering, inputs the sorted dense vectors in time sequence to TCN for prediction, uses sigmoid and tanh activation functions and expands causal convolution to fit parameters, establishes a model, and finally decodes to obtain the ultra-short-term load prediction result.

[0009] The prediction method comprises the following steps:

[0010] Step one, data feature selection and extraction, that is, extracting the data related to the main factors affecting short-term load through the load monitoring system to form the input data for inputting the DTCN model;

[0011] Step two, processing the data by the DCN model module to obtain a dense vector and generate a connection vector x stack ;

[0012] Step three, sending the connection vector into the TCN model module for extracting and predicting the time sequence features;

[0013] Step four, performing the ultra-short-term prediction by the DCN model based on step two, the TCN model of step three and the input data of step one.

[0014] In step one, the data related to the main factors affecting short-term load are historical load data, temperature data and humidity data, the read data are arranged to form the input data for inputting the DTCN model to fit and optimize the mathematical model.

[0015] The historical load data are the annual power data read by the high-voltage side electric meter of the transformer area, and the unit is half an hour.

[0016] In step two, the DCN model starts with an embedding layer and a stacking layer, followed by a cross network and a deep network parallel to the cross network, and then a combination layer;

[0017] Step two comprises the following steps:

[0018] Step S1, processing the input data by the embedding layer of the DCN model, converting the high-dimensional single hot vector in the input data into an embedding vector to reduce the dimension, and the formula is

[0019] X embed，i =w embed，i X i Formula one;

[0020] Wherein X embed，i is an embedding vector, X i is the i-th binary input; and the embedding vector and the normalized continuous features X dense are stacked into a vector:

[0021]

[0022] Finally, x0 is fed back to the network;

[0023] Step S2, splicing the data processed by the cross network and the deep network, wherein the cross network applies explicit feature crossing, and the cross network is composed of cross layers, and each layer can be expressed by the following formula:

[0024]

[0025] wherein X l , X l+1 ∈R d represent the input and output of the cross layer respectively; W l , b l ∈R d represent the weight and offset parameters of the first layer; after each cross layer performs the cross, the input is fed back, and the residual X l+1 -X l is adjusted by a mapping function; in order to overcome the parameter limitation of the single cross network for the DCN, the formula of each layer of the introduced DNN network is as follows:

[0026] h l+1 =f(W l h l +b l ) Formula Four

[0027] wherein h l ∈R nl , h l+1 ∈R nl+1 are the lth hidden layer and the l+1th hidden layer respectively; is the parameter of the lth deep layer.

[0028] In the combination layer, the outputs of the cross network and the deep network are spliced to generate a connection vector x stack , and then x stack is stacked in time sequence and sent to the TCN to continue extracting and predicting the time sequence features.

[0029] The TCN model in step three is composed of dilated, causal 1D convolution layers with the same input and output length;

[0030] Step three includes the following steps:

[0031] Step A1, using the method of expanding causal convolution to process the data in the combination layer of the TCN model, so that the prediction p(x t+1 |x1,..., x t ) output by the model at time t does not depend on any data x t+1 of future time; x t+2 ,..., x T can increase the receptive field by several orders of magnitude without significantly increasing the cost;

[0032] Step A2, using the gate unit as an activation function, using sigmoid and tanh function,

[0033] z=tanh(W f，k *x)⊙σ(W g，k *x) Formula five;

[0034] Where * is convolution, is dot product, sigma is sigmoid function, k is layer index, f and g are filter and gate respectively, W is a learnable convolution filter;

[0035] Step A3, using residual and parameterized skip to speed up convergence; the method is: each node in each layer adds the original value and the value through the activation function to the next layer, and the 1x1 convolution kernel is used to realize the operation of reducing the channel number; then the results of each hidden layer after the activation function are added and a series of operations are performed to pass to the output layer.

[0036] Step A4, output the next output prediction under the TCN model Softmax layer, and optimize to maximize the possibility of data changing with parameters; by adjusting the hyperparameters to measure whether the model is overfitting or underfitting, the model building is completed.

[0037] In step four, the DCN model based on step two and the TCN model of step three are matched to form a DTCN model, real-time load data is collected through the high-voltage side meter of the transformer area, and meteorological data collected through the meteorological monitoring system is used to predict the load of the next day, and the super short-term load change result is obtained.

[0038] The present application proposes a prediction method based on DTCN for data feature extraction and learning, aiming at the problems of strong particularity, low precision and difficulty in efficiently dealing with variable influencing factors of traditional short-term power load prediction method, the present application uses the influence of climate and other characteristics on load change to combine it with load data, and actively extracts effective feature information through DTCN data fitting, automatically extracts features without manual adjustment, solves the problem of difficult to efficiently deal with the influence of complex and variable characteristics on load prediction, and effectively captures the influence of long-term climate characteristics on current load fluctuation, realizes super short-term prediction of load, and has low data capacity requirement, fast prediction speed and high precision.

[0039] The present application has high prediction accuracy and generalization ability in actual system load prediction, is more suitable for power prediction in actual environment compared with other short-term power prediction methods, and can provide reliable energy management for energy management system.

[0040] This invention is a novel method for ultra-short-term forecasting of power load through data fitting. By extracting data from load and meteorological monitoring systems, and fitting the data using the DTCN model, the method obtains load changes in the ultra-short term, thereby balancing supply and demand in a timely manner and maintaining the safe and stable supply of electricity.

[0041] This invention addresses the problems of traditional power load forecasting methods (such as ARMA and SVM) being unable to cope with the impact of variable external factors (such as climate and social activities) on power load changes and having low prediction accuracy. It automatically performs feature engineering on load and climate characteristics through DCN and TCN models respectively, extracts effective features, and establishes mathematical models of the input and output of DTCN, thereby predicting short-term load changes and solving the problems of traditional technologies. Attached Figure Description

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0043] Appendix Figure 1 This is a schematic diagram of the load forecasting process of the present invention;

[0044] Appendix Figure 2 This is a schematic diagram of the model framework of the DTCN of the present invention;

[0045] Appendix Figure 3 This is a diagram comparing the continuous load forecast results of different models, using data from December 14, 2020 as an example.

[0046] Appendix Figure 4 This is a diagram comparing load forecasts from different models, with the forecast point set 12 hours in advance.

[0047] Appendix Figure 5 This is a diagram comparing the prediction models of each model when the predicted data is for the day of 2018-09-11.

[0048] Appendix Figure 6 This is another schematic diagram comparing the prediction models of each model. Detailed Implementation

[0049] As shown in the figure, the power load prediction method based on the Deep Cross Temporal Convolutional Network (DTCN) model is used for ultra-short-term power load prediction. The prediction method utilizes continuous and discrete feature vectors in the dataset, extracts data features through automatic feature engineering using DCN, stacks the sorted dense vectors in time order, and inputs them into TCN for prediction. By using sigmoid and tanh activation functions and expanding causal convolution, the parameters are fitted, the model is established, and finally the ultra-short-term load prediction results are decoded to obtain the prediction results.

[0050] The prediction method includes the following steps;

[0051] Step one, data feature selection and extraction, that is, extracting the main factor related data affecting short-term load through the load monitoring system to form the input data for inputting the DTCN model;

[0052] Step two, processing the data by the DCN model module to obtain a dense vector and generate a connection vector x stack ;

[0053] Step three, sending the connection vector into the TCN model module for extracting and predicting the time sequence features;

[0054] Step four, performing the ultra-short-term prediction by the DCN model based on step two, the TCN model of step three and the input data of step one.

[0055] In step one, the main factor related data affecting the short-term load is historical load data, temperature data and humidity data, the read data is arranged to form the input data for inputting the DTCN model to fit and optimize the mathematical model.

[0056] The historical load data is the previous year's power data read by the high-voltage side meter of the transformer area, with a measurement unit of every half hour.

[0057] In step two, the DCN model starts with an embedding layer and a stacking layer, followed by a cross network and a deep network parallel to the cross network, and then a combination layer;

[0058] Step two includes the following steps.

[0059] Step S1, processing the input data by the embedding layer of the DCN model, converting the high-dimensional single hot vector in the input data into an embedding vector to reduce the dimension, and the formula is

[0060] X embed,i =W embed,i X i Formula one;

[0061] Where X embed,i is the embedding vector, X i is the i-th binary input; and the embedding vector and the normalized continuous feature X dense are stacked into a vector:

[0062]

[0063] Finally, X0 is fed back to the network;

[0064] Step S2, splicing the data processed by the cross network and the deep network, wherein the cross network applies explicit feature crossing, and the cross network is composed of cross layers, and each layer can be expressed by the following formula:

[0065]

[0066] where X l , X l+1 ∈R d represent the input and output of the cross layer respectively; W l , b l ∈R d represent the weight and bias parameters of the first layer; after each cross layer is executed, the input is fed back and the residual X l+1 -X l is adjusted by the mapping function; in order to overcome the parameter limitation of the single cross network for the DCN, the DNN network introduced by the present application has the following formula for each layer:

[0067] h l+1 = f(W l h l +b l ) Formula Four;

[0068] where h l ∈R nl , h l+1 ∈R nl+1 are the lth hidden layer and the l+1th hidden layer respectively; is the parameter of the lth deep layer.

[0069] The outputs of the cross network and the deep network are spliced in the combination layer to generate a connection vector x stack , and x stack is stacked in time sequence and then sent to the TCN to continue extracting and predicting the time sequence features.

[0070] The TCN model in step three is composed of dilated, causal 1D convolution layers with the same input and output length;

[0071] Step three includes the following steps:

[0072] Step A1, the data in the combination layer of the TCN model is processed by using the method of expanding the causal convolution, so that the prediction p(x t+1 |x1,..., x t ) output by the model at time t is not dependent on any future data x t+1 ; x t+2 ,..., x T can increase the receptive field by several orders of magnitude without significantly increasing the cost;

[0073] Step A2, the gate unit is used as the activation function, and the sigmoid and tanh functions are used,

[0074] z = tanh(wf，k X) * sigma(W g，k X) * formula five;

[0075] Wherein * is convolution, is dot product, sigma is sigmoid function, k is layer index, f and g are filter and gate respectively, W is learnable convolution filter;

[0076] Step A3, use residual and parameterized skip to speed up convergence, the method is: each node in each layer adds the original value and the value through the activation function to the next layer, wherein the 1x1 convolution kernel is used to realize the operation of reducing the number of channels, then the results of each hidden layer after the activation function are added and a series of operations are performed to pass to the output layer.

[0077] Step A4, output the next output prediction under the TCN model Softmax layer, and optimize to maximize the possibility of data changing with parameters, measure whether the model is overfitting or underfitting by adjusting the hyperparameters, and complete the model building.

[0078] In step four, the DCN model based on step two and the TCN model of step three are matched to form a DTCN model, real-time load data is collected through the high-voltage side meter of the transformer area, and meteorological data collected through the meteorological monitoring system is used to predict the load of the next day, and the ultra-short-term load change result is obtained.

[0079] The example is aimed at the problems that the traditional power load prediction method (such as ARMA, SVM, etc.) cannot cope with the influence of variable external factors (such as climate, social activities, etc.) on power load change and has low prediction accuracy, respectively, through the DCN and TCN models, the characteristics of load and climate are automatically subjected to feature engineering, effective features are extracted, the input and output mathematical model of DTCN is established, and thus the load change in the short term is predicted.

[0080] As shown in Figures 3 to 6 The application has higher prediction accuracy and generalization ability in actual system load prediction, is more suitable for power prediction in actual environment compared with other short-term power prediction methods, and can provide reliable energy management for the energy management system.

Claims

1. A power load forecasting method based on the Deep Cross-Temporal Convolutional Network (DTCN) model, used for ultra-short-term power load forecasting, characterized by: The prediction method utilizes continuous and discrete feature vectors in the dataset, extracts data features through automatic feature engineering using DCN, stacks the sorted dense vectors in chronological order and inputs them into TCN for prediction, fits parameters and builds a model by using sigmoid and tanh activation functions and enlarged causal convolution, and finally decodes to obtain the ultra-short-term prediction results of the load. The prediction method includes the following steps; Step 1: Data feature selection and extraction, which involves extracting relevant data on the main factors affecting short-term load through the load monitoring system to form input data for the DTCN model; Step 2: Process the data using the DCN model module to obtain dense vectors and generate connection vectors. ; Step 3: Input the connection vector into the TCN model module for temporal feature extraction and prediction; Step 4: Using the DCN model from Step 2 and the TCN model from Step 3, perform ultra-short-term predictions using the input data from Step 1. In step two, the DCN model begins with an embedding layer and a stacking layer, followed by a cross network and a parallel deep network, and then a combination layer. Step two includes the following steps; Step S1: Process the input data through the embedding layer of the DCN model, converting the high-dimensional single hotspot vectors in the input data into embedding vectors, thereby reducing the dimensionality. The formula is as follows: ； Step S2: Concatenate the data processed by the cross-network and the deep network. The cross-network uses explicit feature crossing and consists of cross layers. Each layer can be expressed by the following formula: ； The outputs of the cross-network and the deep network are concatenated in the combination layer to generate a connection vector. , Then After being stacked in chronological order, they are fed into the TCN for further extraction and prediction of temporal features; The TCN model in step three consists of dilated, causal 1D convolutional layers with the same input and output lengths; Step 3 Includes the following steps; Step A1: Process the data in the combined layer of the TCN model using the method of enlarged causal convolution, so that the model... ； Where * represents convolution, For dot product, is the sigmoid function, k is the layer index, f and g are the filter and gate respectively, and W is the learnable convolutional filter; Step A3: Use residuals and parameterized skip steps to speed up convergence. The method is as follows: Each node in the layer adds the original value and the value after passing through the activation function and then passes it to the next layer. The 1x1 convolution kernel is used to reduce the number of channels. Then, the results after passing through the activation function of each hidden layer are added together, a series of operations are performed, and then the results are passed to the output layer. Step A4: Output the next output prediction in the Softmax layer of the TCN model and optimize it to maximize the possibility of data changing with parameters; measure whether the model is overfitting or underfitting by adjusting hyperparameters to complete the model building; In step one, the relevant data of the main factors affecting short-term load are historical load data, temperature data, and humidity data. The obtained data is organized to form the input data for inputting into the DTCN model in order to fit and optimize the mathematical model.

2. The power load forecasting method based on the Deep Cross-Temporal Convolutional Network (DTCN) model according to claim 1, characterized in that: Historical load data is the electricity consumption data from previous years retrieved from the high-voltage side meters of the transformer area, measured in half-hour increments.

3. The power load forecasting method based on the Deep Cross-Temporal Convolutional Network (DTCN) model according to claim 1, characterized in that: In step four, the DCN model based on step two and the TCN model based on step three are combined to form the DTCN model. Load data is collected in real time by the high-voltage side meters of the distribution area, and meteorological data is collected by the meteorological monitoring system to predict the load for the next day and obtain the ultra-short-term load change results.

Images