CN-121980490-A - Sequential prediction method, system, equipment and medium based on CEEMDAN enhanced TCN and QLSTM fusion framework

CN121980490ACN 121980490 ACN121980490 ACN 121980490ACN-121980490-A

Abstract

The invention is suitable for the field of power load prediction, and discloses a time sequence prediction method, a system, equipment and a medium based on CEEMDAN enhanced TCN and QLSTM fusion framework, wherein the method comprises the steps of obtaining original load data, and decomposing by adopting a fully-integrated empirical mode decomposition CEEMDAN method of self-adaptive noise to obtain a high-frequency component sequence and a low-frequency component sequence; the method comprises the steps of constructing a hybrid prediction model, wherein the model comprises a parallel time domain convolution network TCN module and a quantum long and short time memory network QLSTM module, predicting the high-frequency component sequence by adopting the TCN module to obtain a high-frequency prediction result, predicting the low-frequency component sequence by adopting the QLSTM module to obtain a low-frequency prediction result, and fusing the high-frequency prediction result and the low-frequency prediction result to obtain a final prediction result. By combining QLSTM with TCN, the invention realizes high-efficiency prediction of low-frequency and high-frequency information, reduces the demand of computing resources, and reduces the energy consumption and cost in the power load prediction process.

Inventors

GAO ZHENGHAO
WU XIN
Deng Yuedan
XIAO YANHONG
XU KUI

Assignees

贵州电网有限责任公司

Dates

Publication Date: 20260505
Application Date: 20251218

Claims (10)

1. A method for timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework, comprising: the method comprises the steps of obtaining original load data, and decomposing by adopting a full set empirical mode decomposition CEEMDAN method of self-adaptive noise to obtain a high-frequency component sequence and a low-frequency component sequence; Constructing a hybrid prediction model, wherein the model comprises a parallel time domain convolution network TCN module and a quantum long and short time memory network QLSTM module; Predicting the high-frequency component sequence by adopting a TCN module to obtain a high-frequency prediction result, and predicting the low-frequency component sequence by adopting a QLSTM module to obtain a low-frequency prediction result; and fusing the high-frequency prediction result and the low-frequency prediction result to obtain a final prediction result.
2. The method for timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework according to claim 1, wherein the steps of obtaining raw load data and decomposing by using a full set empirical mode decomposition CEEMDAN method of adaptive noise to obtain a high frequency component sequence and a low frequency component sequence include: Adding paired self-adaptive noise into the original load data, and performing empirical mode decomposition on the load data after noise addition to obtain a plurality of intrinsic mode function IMF components and residual items; And constructing a plurality of subsequences by the IMF components and residual terms of the plurality of eigen mode functions, and reconstructing the subsequences to obtain a high-frequency component sequence and a low-frequency component sequence.
3. The method for predicting the time sequence based on CEEMDAN enhanced TCN and QLSTM fusion framework according to claim 2, wherein the quantum long and short time memory network QLSTM module comprises: the quantum long and short time memory network QLSTM module comprises a plurality of stacked quantum long and short time memory network QLSTM units; A single QLSTM unit utilizes multiple variable component sub-circuits VQC to construct the forget gate, input gate and output gate.
4. A method for timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework as in claim 3, wherein said variable component sub-circuit VQC is built up of a feature coding layer, a trainable parameter and entanglement layer and a measurement layer, comprising: Performing unitary transformation coding on the multi-feature representation through a feature coding layer; evolution of the encoded quantum state is carried out by adopting a plurality of layers of trainable parameters and entanglement layers; and measuring the evolved quantum state through a measuring layer to obtain the output characteristics of the variable component sub-circuit VQC.
5. The method of timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework of claim 4, wherein the single QLSTM unit processing logic comprises: splicing the hidden state of the previous moment with the input vector of the current moment to obtain a connection input vector; processing the concatenated input vector by a first VQC and Activating a function to obtain a forgetting door vector, and screening the unit state at the last moment; processing the concatenated input vector by a second VQC and Activating a function to obtain an input gate vector, processing the input vector by a third VQC, obtaining a new unit state candidate through hyperbolic tangent Tanh activating function, and updating the new unit state candidate based on the input gate vector to obtain a current unit state; processing the concatenated input vector by a fourth VQC and Activating a function to obtain an output gate vector, and combining the output gate vector with the current unit state to obtain a to-be-processed result; And processing the result to be processed by using the fifth VQC and the sixth VQC respectively to obtain the hidden state and the output result at the current moment.
6. The method for timing prediction based on CEEMDAN-enhanced TCN and QLSTM fusion framework of claim 5, wherein the time domain convolutional network TCN module comprises: Receiving a high-frequency component sequence as input data, and processing the input data through a trunk branch and a residual branch which are parallel in a TCN (time domain convolutional network) module; in a trunk branch, carrying out convolution operation on input data by utilizing cavity convolution to extract time sequence characteristics, and sequentially carrying out ReLu activation and Dropout processing on the convolved data to obtain a first characteristic result; in the residual branch, carrying out channel adjustment on input data by utilizing point-by-point convolution to obtain a second characteristic result; and adding the first characteristic result and the second characteristic result by elements to obtain a high-frequency prediction result.
7. The method for timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework of claim 6, wherein the hybrid prediction model further performs parameter optimization using a levenberg-marquardt algorithm, comprising: calculating an error vector between the final prediction result and the real load value; calculating a jacobian matrix of model parameters based on the error vector; constructing a linear equation set containing a jacobian matrix and damping factors, and solving to obtain a parameter updating step length; and dynamically adjusting the damping factor according to the error change trend after parameter updating, and updating the VQC and TCN parameters in the mixed prediction model until convergence conditions are met.
8. A timing prediction system based on CEEMDAN enhanced TCN and QLSTM fusion framework, applying the method of any of claims 1-7, comprising: The data decomposition module is used for acquiring original load data, and decomposing by adopting a full set empirical mode decomposition CEEMDAN method of self-adaptive noise to obtain a high-frequency component sequence and a low-frequency component sequence; the model construction module is used for constructing a hybrid prediction model, and the model comprises a parallel time domain convolution network TCN module and a quantum long and short time memory network QLSTM module; The prediction module is used for predicting the high-frequency component sequence by adopting a TCN module to obtain a high-frequency prediction result, and predicting the low-frequency component sequence by adopting a QLSTM module to obtain a low-frequency prediction result; And the fusion module is used for fusing the high-frequency prediction result and the low-frequency prediction result to obtain a final prediction result.
9. An electronic device, comprising: A memory and a processor; the memory is configured to store computer executable instructions that, when executed by a processor, implement the steps of the method for timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework of any one of claims 1 to 7.
10. A computer readable storage medium having stored thereon computer executable instructions which when executed by a processor implement the steps of a method of timing prediction based on CEEMDAN enhanced TCN and QLSTM fusion framework of any one of claims 1 to 7.

Description

Sequential prediction method, system, equipment and medium based on CEEMDAN enhanced TCN and QLSTM fusion framework Technical Field The invention relates to the field of power load prediction, in particular to a time sequence prediction method, a system, equipment and a medium based on CEEMDAN enhanced TCN and QLSTM fusion frames. Background Power load prediction is a key component in power system planning and operation management, solving key problems such as power supply and demand balance, power grid planning, economic dispatch, energy cost management, system reliability and stability. The importance of the economic feasibility, the reliability and the environmental sustainability of the power system is ensured. With advances in technology, load prediction methods have evolved to meet the dynamic demands of modern power systems. Conventional LSTM-TCN mixed modes exhibit significant capabilities in processing multi-temporal scale features. However, as the depth and complexity of the model increases, the computational cost increases exponentially. Therefore, the invention contemplates a QLSTM-TCN hybrid module, which ingeniously utilizes the equivalent sub-characteristics of superposition, entanglement and quantum parallelism. The module seamlessly integrates QLSTM and TCN, and fully utilizes the advantages of the two components. Inspired by these advances, innovations introduced a new neural network architecture called CEEMADAN (full set empirical mode decomposition with adaptive noise) -QLSTM (quantum long short-term memory network) -TCN (time domain convolutional network). QLSTM has demonstrated its potential in sequence prediction and text classification tasks, emphasizing the feasibility of combining quantum computing with neural networks, significantly reducing model complexity and speeding up convergence. By utilizing quantum mechanics and classical computing techniques, the aim is to innovate the field of load prediction, providing a more efficient and scalable solution to cope with complex power system challenges. Disclosure of Invention The invention is provided in view of the problems of complex model, high calculation cost and limited precision existing in the prior art. Therefore, the invention provides a time sequence prediction method, a system, equipment and a medium based on CEEMDAN enhanced TCN and QLSTM fusion frames, which solve the problems that the existing original power load data contains a large amount of randomness and non-stationarity, the direct prediction results in low precision, the traditional decomposition method is easy to have modal aliasing, a single model is difficult to simultaneously consider the feature extraction of short-term severe fluctuation (high frequency) and long-term trend dependence (low frequency), and the time sequence prediction model cannot be optimized by effectively utilizing the parameter advantage of quantum computation when high-dimensional features are processed. In order to solve the technical problems, the invention provides the following technical scheme: in a first aspect, the present invention provides a method for predicting a time sequence based on CEEMDAN enhanced TCN and QLSTM fusion framework, including: the method comprises the steps of obtaining original load data, and decomposing by adopting a full set empirical mode decomposition CEEMDAN method of self-adaptive noise to obtain a high-frequency component sequence and a low-frequency component sequence; Constructing a hybrid prediction model, wherein the model comprises a parallel time domain convolution network TCN module and a quantum long and short time memory network QLSTM module; Predicting the high-frequency component sequence by adopting a TCN module to obtain a high-frequency prediction result, and predicting the low-frequency component sequence by adopting a QLSTM module to obtain a low-frequency prediction result; and fusing the high-frequency prediction result and the low-frequency prediction result to obtain a final prediction result. The invention relates to a time sequence prediction method based on CEEMDAN enhanced TCN and QLSTM fusion frame, which comprises the steps of obtaining original load data, and decomposing by adopting a complete set empirical mode decomposition CEEMDAN method of self-adaptive noise to obtain a high-frequency component sequence and a low-frequency component sequence, wherein the method comprises the following steps: Adding paired self-adaptive noise into the original load data, and performing empirical mode decomposition on the load data after noise addition to obtain a plurality of intrinsic mode function IMF components and residual items; And constructing a plurality of subsequences by the IMF components and residual terms of the plurality of eigen mode functions, and reconstructing the subsequences to obtain a high-frequency component sequence and a low-frequency component sequence. As a preferable scheme of the time sequence prediction meth