Machine-learning-based capacity prediction and construction parameter optimization for energy storage salt caverns

The construction design and control of energy storage salt caverns is the key to ensure their long-term storage capacity and operational safety. Current experimental and numerical design/optimizing methods are time-consuming and rely heavily on engineering experience. This paper proposes a machinelearning-based method for the rapid capacity prediction and construction parameter optimization of energy storage salt caverns. We propose a data generation method that uses 1253 sets of random construction parameters as input. The resulting capacity/efficiency-concerned effective volume (V) and maximum radius (rmax) obtained by our numerical program are the output. A back-propagation artificial neural network model for salt cavern construction prediction (BPANN-SCCP) is trained on the dataset. The cross-validated mean absolute percentage error (MAPE) of the BPANN-SCCP predicted Vis 1.838%, that of the predicted rmax is 3.144%. This accuracy meets the engineering design requirements, and the prediction efficiency is improved by about 6 x 107 times. Using this model, a design parameter optimization method is devised to optimize 3 sets of design parameters from a million random ones. The resulting caverns are regular in shape with larger capacity ratio than 3 field caverns in Jintan Salt Cavern Gas Storage, verifying the reliability of the proposed optimization method. (c) 2022 Elsevier Ltd. All rights reserved.