Mine mining power consumption prediction system and method based on deep learning

Abstract

The invention relates to a mining electric power consumption prediction system based on deep learning. The mining electric power consumption prediction system comprises a time convolution network module and an Informer module which are parallel and are based on a time attention mechanism. The invention further relates to a mining power consumption prediction method based on deep learning. Modules with different advantages are improved and combined to construct a parallel combination model suitable for mining power consumption prediction. According to the method, a smaller mean square error and a smaller mean absolute error can be obtained, so that compared with an existing method, the method has higher prediction performance, mine enterprises can more accurately and reasonably plan future electricity utilization, and the utilization rate of electric energy is increased.

Description

Technical Field

[0001] The present invention belongs to the technical field of power energy consumption prediction, and specifically relates to a system and method for predicting power energy consumption in mining based on deep learning. Background Art

[0002] Mining is an indispensable part of modern industrialized production. As one of the important energy sources for mining, electricity consumption is directly related to the production cost and economic benefits of enterprises. At present, with the reform of the power market and the optimization of the energy structure, mining enterprises are facing the pressure of rising electricity costs. In addition, with the continuous consumption of national energy reserves and the increasing prominence of environmental problems, how to reasonably plan the consumption of electricity in the mining process has become a very urgent issue. Therefore, predicting and studying the consumption of electricity in the mining process will help improve the energy efficiency of enterprises, reduce the production costs of enterprises, and more accurately control and plan the use of mining electricity, thereby promoting the sustainable development of enterprises.

[0003] Deep learning (DL) is a branch of machine learning (ML). Its main feature is to use a neural network model with multiple hidden layers to learn complex nonlinear mapping relationships between input and output. These networks can automatically extract features from data without manually designing feature extraction rules. Deep learning has achieved good results in image recognition, speech recognition, natural language processing and other fields.

[0004] Temporal Convolutional Network (TCN) is a new type of deep learning model for processing sequence data. It combines the advantages of convolutional neural network (CNN) and recurrent neural network (RNN), and is mainly used for time series prediction and sequence modeling tasks. TCN uses causal convolution to ensure that the model can only rely on the current and previous time steps when predicting the current time step, avoiding future information leakage. At the same time, by expanding convolution, TCN can increase the receptive field while maintaining a small amount of computation, allowing the model to capture long-term dependencies. In addition, it also introduces residual connections to accelerate training and improve model performance.

[0005] The Informer architecture is a special variant of the Transformer architecture, which is proposed to improve the Transformer's ability to process long-span time series data. The traditional Transformer's self-attention mechanism has a high computational complexity, which dramatically increases the demand for computing resources when processing long sequences. Secondly, it has a limit on the length of the input sequence, which easily leads to information loss. The Informer architecture has made a series of important improvements on the basis of the Transformer, aiming to improve the efficiency and performance of long sequence data processing. First, it introduces a sparse self-attention mechanism, which significantly reduces the computational complexity by selectively focusing on key time steps, making it more efficient when processing long sequences. In addition, the Informer adopts an information selection module to dynamically select the historical information that is most helpful for prediction, thereby improving accuracy. The application of the formula autoregressive prediction method enables the model to better capture the dynamic changes of the sequence, especially in the time series prediction task. Through a multi-level prediction mechanism, the Informer can extract features at different levels and enhance the ability to model complex time series patterns. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a system and method for predicting power consumption in mining based on deep learning, so as to solve the problem of low prediction accuracy of the existing power consumption prediction technology.

[0007] The present invention solves the technical problem by the following technical solutions:

[0008] A deep learning-based mining power consumption prediction system includes a parallel temporal convolutional network module and an informer module based on a temporal attention mechanism; the temporal convolutional network module is used to extract local features of historical data, and the informer module is used to extract global features of historical data and predict power consumption in the future.

[0009] Moreover, the temporal convolutional network module includes a plurality of residual modules and an average pooling layer connected in sequence; each residual module includes a convolution unit A, a temporal attention mechanism module A, a convolution unit B and a temporal attention mechanism module B connected in sequence; the convolution unit A and the convolution unit B both include an expansion causal convolution layer, a weight normalization layer, an activation function layer and a regularization layer connected in sequence.

[0010] Moreover, each of the residual modules also includes a convolution layer A. The data input to the convolution unit A is also input to the convolution layer A. The output of the convolution layer A is fused with the output of the temporal attention mechanism module B and then enters the average pooling layer, which is then used as the output of the residual module after pooling.

[0011] Moreover, the temporal attention mechanism module A and the temporal attention mechanism module B both include a convolution layer B and a sigmoid activation function layer; the convolution layer B convolves the input vector y1, generates a weight coefficient W using the sigmoid function, and then multiplies the weight coefficient by the original input vector y1 to form a new output vector y2; the mathematical expressions of the temporal attention mechanism module A and the temporal attention mechanism module B are as follows:

[0012] W = σ(Conv1D(y1));

[0013]

[0014] Where: σ represents the sigmoid function;

[0015] Conv1D represents one-dimensional convolution;

[0016] o represents the Hadamard product.

[0017] Moreover, the Informer module includes a plurality of encoders and decoders connected in sequence.

[0018] Moreover, the encoder includes a multi-head sparse self-attention layer A, a distillation layer A, a multi-head sparse self-attention layer B and a distillation layer B connected in sequence, and the mathematical expressions of the multi-head sparse self-attention layer A and the multi-head sparse self-attention layer B are as follows:

[0019]

[0020] in: is a sparse matrix of the same size as q, but containing only the largest L in M K LqCy Q The space and time complexity of a query is only O(L ln L) when calculating, thus reducing the space complexity and time complexity;

[0021] The mathematical expressions of distillation layer A and distillation layer B are as follows:

[0022]

[0023] in: For the attention block, after each convolutional layer, the distillation operation generates a maximum pooling layer with a stride of 2. Downsampling is performed to reduce the computational space complexity to O((2-λ)L log L), where λ is a constant.

[0024] Moreover, the decoder includes a masked multi-head probabilistic sparse self-attention layer A and a multi-head attention layer A connected in sequence, and finally connected to a fully connected layer A to output the prediction result.

[0025] A method for predicting mining power consumption based on deep learning, using the prediction system, the steps of the method are:

[0026] S1, construct a combined model including a temporal convolutional network module based on the temporal attention mechanism and an Informer module;

[0027] S2, collects historical data of power consumption of a mine, pre-processes the collected raw data to match the input format of the combined model and compiles it into a data set; divides the data set into a training set, a validation set and a test set;

[0028] S3, input the training set into the combined model for training; and use the validation set to adjust the model parameters;

[0029] S4, input the test set into the trained combined model for prediction to obtain the final prediction value;

[0030] S5, the prediction performance of the combined model was evaluated using mean square error and mean absolute error.

[0031] The positive effects that the present invention can produce are:

[0032] The present invention improves and combines modules with different advantages to construct a parallel combination model suitable for predicting power consumption in mining. The method of the present invention can obtain a smaller mean square error and mean absolute error, indicating that the present invention has stronger prediction performance than the existing methods, and can enable mining enterprises to more accurately plan future electricity consumption and improve the utilization rate of power energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the improved temporal convolutional network module based on the temporal attention mechanism of the present invention;

[0034] Figure 2 This is a schematic diagram of the structure of the Informer module based on the sparse attention mechanism of the present invention;

[0035] Figure 3 It is the process and structure diagram of the present invention. DETAILED DESCRIPTION

[0036] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0037] like Figures 1 to 3As shown, a mining power consumption prediction system based on deep learning has the following innovations: it includes a parallel temporal convolutional network module and an Informer module based on a temporal attention mechanism; the temporal convolutional network module is used to extract local features of historical data, and the Informer module is used to extract global features of historical data and predict power consumption in the future.

[0038] Moreover, the temporal convolutional network module includes a plurality of residual modules and an average pooling layer connected in sequence; each residual module includes a convolution unit A, a temporal attention mechanism module A, a convolution unit B and a temporal attention mechanism module B connected in sequence; the convolution unit A and the convolution unit B both include an expansion causal convolution layer, a weight normalization layer, an activation function layer and a regularization layer connected in sequence.

[0039] Moreover, each of the residual modules also includes a convolution layer A. The data input to the convolution unit A is also input to the convolution layer A. The output of the convolution layer A is fused with the output of the temporal attention mechanism module B and then enters the average pooling layer, which is then used as the output of the residual module after pooling.

[0040] Moreover, the temporal attention mechanism module A and the temporal attention mechanism module B both include a convolution layer B and a sigmoid activation function layer; the convolution layer B convolves the input vector y1, generates a weight coefficient W using the sigmoid function, and then multiplies the weight coefficient by the original input vector y1 to form a new output vector y2; the mathematical expressions of the temporal attention mechanism module A and the temporal attention mechanism module B are as follows:

[0041] W = σ(Conv1D(y1));

[0042]

[0043] Where: σ represents the sigmoid function;

[0044] Conv1D represents one-dimensional convolution;

[0045] o represents the Hadamard product.

[0046] Moreover, the Informer module includes a plurality of encoders and decoders connected in sequence.

[0047] Moreover, the encoder includes a multi-head sparse self-attention layer A, a distillation layer A, a multi-head sparse self-attention layer B and a distillation layer B connected in sequence, and the mathematical expressions of the multi-head sparse self-attention layer A and the multi-head sparse self-attention layer B are as follows:

[0048]

[0049] in: is a sparse matrix of the same size as q, but containing only the largest L in M K LqCy Q The space and time complexity of a query is only O(L ln L) when calculating, thus reducing the space complexity and time complexity;

[0050] The mathematical expressions of distillation layer A and distillation layer B are as follows:

[0051]

[0052] in: For the attention block, after each convolutional layer, the distillation operation generates a maximum pooling layer with a stride of 2. Downsampling is performed to reduce the computational space complexity to O((2-λ)L log L), where λ is a constant.

[0053] Moreover, the decoder includes a masked multi-head probabilistic sparse self-attention layer A and a multi-head attention layer A connected in sequence, and finally connected to a fully connected layer A to output the prediction result.

[0054] A method for predicting mining power consumption based on deep learning, the innovation of which is: using the prediction system, the steps of the method are:

[0055] S1, construct a combined model including a temporal convolutional network module based on the temporal attention mechanism and an Informer module;

[0056] S2, collects historical data of power consumption of a mine, pre-processes the collected raw data to match the input format of the combined model and compiles it into a data set; divides the data set into a training set, a validation set and a test set;

[0057] S3, input the training set into the combined model for training; and use the validation set to adjust the model parameters;

[0058] S4, input the test set into the trained combined model for prediction to obtain the final prediction value;

[0059] S5, the prediction performance of the combined model was evaluated using mean square error and mean absolute error.

[0060] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A mining power consumption prediction system based on deep learning, characterized by: It includes a parallel temporal convolutional network module and an Informer module based on a temporal attention mechanism; the temporal convolutional network module is used to extract local features of historical data, and the Informer module is used to extract global features of historical data and predict power consumption in the future.

2. The mining power consumption prediction system based on deep learning according to claim 1 is characterized in that: The temporal convolutional network module includes a plurality of residual modules and an average pooling layer connected in sequence; each residual module includes a convolution unit A, a temporal attention mechanism module A, a convolution unit B and a temporal attention mechanism module B connected in sequence; the convolution unit A and the convolution unit B both include an expansion causal convolution layer, a weight normalization layer, an activation function layer and a regularization layer connected in sequence.

3. The mining power consumption prediction system based on deep learning according to claim 2 is characterized in that: Each of the residual modules also includes a convolution layer A. The data input to the convolution unit A is also input to the convolution layer A. The output of the convolution layer A is fused with the output of the temporal attention mechanism module B and then enters the average pooling layer, which is then used as the output of the residual module after pooling.

4. The mining power consumption prediction system based on deep learning according to claim 2 is characterized in that: The temporal attention mechanism module A and the temporal attention mechanism module B each include a convolution layer B and a sigmoid activation function layer; the convolution layer B convolves the input vector y1, generates a weight coefficient W using the sigmoid function, and then multiplies the weight coefficient by the original input vector y1 to form a new output vector y2; the mathematical expressions of the temporal attention mechanism module A and the temporal attention mechanism module B are as follows: W = σ(Conv1D(y1)); Where: σ represents the sigmoid function; Conv1D represents one-dimensional convolution; o represents the Hadamard product.

5. The mining power consumption prediction system based on deep learning according to claim 1 is characterized by: The Informer module includes a plurality of encoders and decoders connected in sequence.

6. The mining power consumption prediction system based on deep learning according to claim 5 is characterized by: The encoder includes a multi-head sparse self-attention layer A, a distillation layer A, a multi-head sparse self-attention layer B and a distillation layer B connected in sequence. The mathematical expressions of the multi-head sparse self-attention layer A and the multi-head sparse self-attention layer B are as follows: in: is a sparse matrix of the same size as q, but containing only the largest L in M K l Q The space and time complexity of a query is only O(L ln L) when calculating, thus reducing the space complexity and time complexity; The mathematical expressions of distillation layer A and distillation layer B are as follows: in: For the attention block, after each convolutional layer, the distillation operation generates a maximum pooling layer with a stride of 2. Downsampling is performed to reduce the computational space complexity to O((2-λ)L log L), where λ is a constant.

7. The mining power consumption prediction system based on deep learning according to claim 5 is characterized by: The decoder includes a masked multi-head probabilistic sparse self-attention layer A and a multi-head attention layer A connected in sequence, and finally connected to a fully connected layer A to output the prediction result.

8. A method for predicting mining power consumption based on deep learning, characterized in that: Using the prediction system according to any one of claims 1 to 7, the steps of the method are: S1, construct a combined model including a temporal convolutional network module based on the temporal attention mechanism and an Informer module; S2, collects historical data of power consumption of a mine, pre-processes the collected raw data to match the input format of the combined model and compiles it into a data set; divides the data set into a training set, a validation set and a test set; S3, input the training set into the combined model for training; and use the validation set to adjust the model parameters; S4, input the test set into the trained combined model for prediction to obtain the final prediction value; S5, the prediction performance of the combined model was evaluated using mean square error and mean absolute error.

Images