Accurate multi-horizon solar power forecasting is vital for modern power distribution networks, yet standard deep learning models face limitations due to weather stochasticity and prohibitive computational requirements on resource-constrained edge devices. To address these issues, this paper proposes a Physics-Informed Similar-Day Data Augmentation (PI-SDDA) framework coupled with a Pruning-Optimized Temporal Convolutional Network (POTCN). The PI-SDDA strategy integrates multi-variable meteorological dynamics with candidate cluster size optimization to capture diurnal thermal periodicity and transient weather variations. To enable lightweight deployment, POTCN incorporates Batch Normalization scaling factors for channel-level sparsification within dilated causal residual blocks, significantly reducing parameter redundancy while preserving temporal receptive fields. Experimental results demonstrate superior forecasting performance across fifteen-minute, one-hour, and two-hour prediction horizons under standard statistical metrics compared with existing benchmarks. Edge hardware deployment tests on an embedded edge computing platform further verify low inference latency and minimal power consumption. The proposed framework provides an accurate, resource-efficient solution for real-time edge intelligence in power systems.