PILSTM-PIT: A Multimodal Model for Forecasting Fluctuating Soil Temperature Measurement Data

Accurate estimation and prediction of soil temperature in in situ thermal remediation can help to adopt qualified strategies and directly determine soil remediation efficiency. Due to the complex and diverse structure of soil gaps, the air flow inside drilling causes severe temperature fluctuations, rendering traditional prediction methods ineffective. To address this issue, this article proposes PILSTM-PIT, a novel multimodal data hybrid prediction model that combines physical model data and sensor measurement data to predict temperature changes at monitoring points. In particular, the model integrates physical constraints to reduce the overall bias of prediction results when sensor data are small, introduces Pearson multihead attention to capture sequence time-step correlation, and tracks dynamic fluctuations. Experimental results demonstrate that the model achieves optimal or near-optimal performance in the 14/16 sensor datasets and demonstrates competitive and, in several cases, superior forecasting performance against Transformer-based baselines in the “96-to-96 h” prediction task on the public traffic and ECL benchmarks, with mean absolute error (MAE) and mean-square error (mse) metrics. This study has opened up a new field that allows for high-precision predictions of complex physical activities in the real world using imperfect model data and low-quality measurement data, providing a high-quality solution for industrial measurement data analysis in the I&M field.