MSWI Multi-Temperature Prediction Based on Patch Time Series Transformer

In view of the substantial amount of solid waste in China and the incomplete combustion of solid waste in the incineration process due to unstable combustion temperatures. This leads to the release of a large amount of harmful gases such as dioxins and carbon monoxide into the atmosphere. To ensure the complete combustion of solid waste, temperature prediction control of the incinerator is one of the key measures. Currently, temperature prediction still relies on obtaining information from individual time steps, which is not ideal for long time series temperature prediction. To address these issues, this paper proposes a temperature prediction method for Municipal Solid Waste Incinerator (MSWI) based on the patch time series transformer. Firstly, each input univariate time series is divided into patches. Secondly, a standard Transformer encoder is employed to map the observed signals to the latent representation space. The patches are mapped to the Transformer latent space using a trainable linear projection matrix, and a learnable additive position encoding matrix is applied to monitor the temporal order of the patches. Additionally, to achieve better prediction results, a convolutional layer is added to the encoder for feature extraction. Finally, a flatten layer with a linear head is used to obtain the prediction result. Experimental results of temperature prediction based on the Patch Time Series Transformer model demonstrate promising outcomes for long time series with multiple temperatures.