Abstract:

In actual industrial production， Informer model will lose a large number of temporal characteristics during feature extraction due to its probabilistic sparsity mechanism. To address this shortcoming of Informer model and meet the demands for prediction speed and efficiency in industrial production， a Convolutional Neural Network （CNN）-enhanced Informer model was proposed. Short-Time Fourier Transform （STFT） was introduced to process sequence to obtain frequency-domain features of the data， thereby further reducing feature loss caused by probabilistic sparse attention mechanism and improving prediction accuracy. On public datasets， ETT （Electricity Transformer Temperature）， ECL （Electricity Consumption Load）， as well as one private dataset， comparative experiments were conducted between the proposed model and four models widely used in the industrial field， including Long Short-Term Memory （LSTM） and AutoRegressive Integrated Moving Average （ARIMA） model. Experimental results show that both Mean Squared Error （MSE） and Mean Absolute Error （MAE） of the proposed model are decreased.

Key words: time series prediction, industrial data, Convolutional Neural Network (CNN), Informer model, Short-Term Fourier Transform (STFT)

摘要：

在实际的工业生产中，Informer模型自身的概率稀疏机制会导致在特征提取时大量时序特性的丢失。为了克服Informer模型的这种缺陷，同时兼顾工业生产中对预测速度和效率的要求，提出一种使用卷积神经网络（CNN）增强的Informer模型。该模型引入短时傅里叶变换（STFT）处理序列获取数据在频域的特征，以进一步减少概率稀疏注意力机制带来的特征丢失，并提高预测准确度。在ETT（Electricity Transformer Temperature）、ECL（Electricity Consumption Load）公开数据集和一个私有数据集上，所提模型与工业领域应用最广泛使用的长短期记忆网络（LSTM）、自回归积分滑动平均（ARIMA）模型等4种模型进行对比实验的结果表明，所提模型的均方误差（MSE）和平均绝对误差（MAE）两项指标均有下降，性能有所提升。

关键词: 时序预测, 工业数据, 卷积神经网络, Informer模型, 短时傅里叶变换