Carpet-shaped periodic phase-change thermal energy storage for building heating: numerical and ANN analysis

Solar-assisted domestic-hot-water systems require compact thermal buffers to improve the temporal match between solar heat availability and building hot-water demand. This study investigates transient melting and solidification in a two-dimensional periodic unit cell comprising broken-cross paraffin passages embedded in a copper matrix. An enthalpy-porosity finite-element model resolves conduction in copper, buoyancy-driven flow in liquid PCM, phase evolution, and energy exchange. Periodic boundary conditions represent the repeating carpet-shaped array, while adjacent water-based heat-transfer-fluid channels are represented by convection at the external copper surfaces. The numerical database comprises 19 parameter configurations, each evaluated during melting and solidification, yielding 38 transient histories and 414,376 time-stamped records. A compact deep neural network (DNN) predicts the domain-averaged melt-volume fraction from operating mode, NP, τ0, Fh and elapsed time. Generalization was evaluated by holding out both complete mode histories of one parameter configuration at a time. The nested whole-design evaluation gives an equal-design/mode, time-weighted RMSE of 0.0207 in the original MVF scale. Increasing NP from 1 to 5 lengthens the times to MVF = 0.9 during melting and MVF = 0.1 during solidification by 219.34% and 165.22%, respectively, while increasing absorbed and released energy by 354.65% and 341.53%. Increasing τ0 from 1 to 4 reduces the corresponding times by 39.01% and 42.69%, with energy reductions of 5.77% and 5.04%. Raising Fh from 1 to 4 shortens the times by 29.69% and 27.46% and produces small energy increases of 2.02% and 2.08%. The DNN response curves describe supported parameter slices.