<p>This paper presents a novel self-healing (SH) framework for Solar Photovoltaic Power (SPVP) forecasting using an Attention-based Peephole connected Gated Recurrent Unit Neural Network (A-PGRU NN). Inspired by the natural healing behaviour of termites to maintain their mounds with the help of coordinated teamwork, the proposed Termite Mount-Inspired Neural Network Self-Healing Improved Algorithm (TMNN-SHA) model autonomously detects and recovers from disruptions in the neural network for the forecasting process with the help of coordinated remaining network portions. It integrates four key mechanisms: the Probabilistic Damage Recovery Mechanism (PDRM) for identifying and compensating missing data, Dynamic State Approximation (DSA) for reconstructing disrupted hidden states, Bayesian Confidence-Based Weight Reallocation (BCBWR) to prioritize reliable learning paths, and the Adaptive Loss Adjustment Strategy (ALAS) to dynamically tune the loss function based on damage severity. These are supported by a Reinforcement Learning (RL)-based feedback loop that optimizes recovery actions over time. The model is evaluated on three datasets: NREL benchmark data, real-time measurements from a 1&#xa0;MW SPVP plant in South India, and synthetically corrupted datasets. Experimental results show that TMNN-SHA outperforms baseline models, including Gated Recurrent Unit Neural Network (GRU NN), Long Short-Term Memory NN (LSTM NN) and Bi-directional LSTM (Bi-LSTM), achieving resilience scores consistently above 0.74 and self-healing latency between 20 and 26&#xa0;ms under varying seasonal and damage conditions without affecting the accuracy of the network. The performance improvements of the proposed model are supported by comparisons to state-of-the-art methods in recent literature. The proposed model offers a robust, adaptive, and intelligent solution for dynamic real-world environments with reliable SPVP forecasting.</p>