Fast thermal-state sensing for solid thermal storage in heavy-oil recovery

Abstract

Renewable-electricity substitution in heavy-oil thermal recovery 
requires rapid knowledge of the thermal state inside high-temperature solid 
thermal storage units, where dense internal sensing is difficult to deploy. This 
study develops a sparse-measurement soft-sensing method that combines a one
dimensional convolutional neural network with a bidirectional long short-term 
memory network (CNN-BiLSTM). A field-calibrated multiphysics heat-transfer 
model was first used to generate temperature data over complete charging and 
discharging processes. The model then used 60-step historical sequences from 
five thermocouple features as inputs. The 1D-CNN extracted local thermal
response features among measurement points, whereas the BiLSTM captured 
thermal inertia and temporal lag within the available historical window. The 
target was the volumetric average temperature, which represents the stored 
thermal inventory and thermal state of charge for steam-supply scheduling and 
heat-supply-capacity assessment. Local maximum temperature should still be 
monitored by independent protection sensors. On a chronological test set, five 
repeated training runs yielded a mean absolute error of 5.95 °C, a root mean 
square error of 7.43 °C and an R² of 0.996, with a 95% confidence interval of 
6.51-8.36 °C for RMSE. Under 5% input noise, the RMSE was 7.56 °C and R² 
remained 0.996. The mean single-sample response time, including normalisation, 
forward inference and inverse normalisation, was approximately 23.94 ms. These 
results indicate that the method can support online state estimation and 
operational monitoring of large solid thermal storage systems within the tested 
operating range.