During the operation of storage tanks, the impact of earthquakes is inevitable, leading to significant fluctuations in the liquid levels within the tanks. To proactively prevent operational accidents caused by such level changes and ensure the long-term stable operation of petrochemical tank farms, a dynamic characteristic-driven predictive strategy was developed to evaluate liquid climbing behavior on storage tank walls under external loads with varying fill levels. A scaled 2D experimental model was used to establish a hydrodynamic analysis system to study sloshing responses under harmonic excitation. A dataset was constructed, incorporating both resonance-induced maximum and steady-state liquid elevations. A multilayer perceptron (MLP) model was trained using 40% ~ 50% of the samples as the training set, with the remaining data used for validation. The model achieved prediction errors of less than 4%, an correlation coefficient value close to 1, root mean square error values below 2, and deviations under 5%. This deep learning-based framework enables real-time risk warnings for circumferential pressure calculations and free-surface dynamics assessments, thereby supporting proactive safety management of tanks under varying operating conditions.