The increasing integration of renewable energy sources in decentralized nano-grid systems presents new challenges in maintaining load balance and ensuring fault resilience due to the intermittent and nonlinear nature of energy flows. In this paper, we introduce a hybrid CNN-GRU architecture integrated with AI that exploits the advantages of both Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRUs) in a time-sensitive energy management context within nano-grid environments. The model is trained on the room temperature in the earliest conversations, using high resolution time-series measurements obtained by smart meters and energy nodes, specifically the parameters of voltage, current, frequency, and state-of-charge ( SoC ). The proposed model outperformed conventional methods like ARIMA, Random Forest and LSTM, standalone CNN and GRU in terms of accuracy and Mean Absolute Percentage Error of forecasting load (3.14%) and fault classification (96.94%) having conducted the large-scale experimentation. In addition, the model attained low inference latency of less than 15 milliseconds and has a fault response time of less than 500 milliseconds, which confirmed that it was applicable in edge deployment. Experimentations in the real world demonstrated an energy consumption reduction of up to 20% in different operating conditions related to peak demand, renewable surplus, and grid failure. In our research, we show that by integrating deep learning with edge-intelligent control it is possible to greatly increase the resilience, adaptability and sustainability of the next generation decentralized power systems. CNN-GRU can be a good use case as the prediction engine in the sense of short-term load balance and real-time fault recovery, and the model may be useful in developing smart energy infrastructure.