This research introduces a model with two distinct stages for accurate forecasting and efficient energy sharing within grid-integrated community networks (GICNs), which combines residential power loads, generators of wind and photovoltaic (PV) power, and energy storage (ES). In the first-stage system model, we employ a machine learning (ML) algorithm for day-ahead supply and demand forecasting to improve the forecasting accuracy. Specifically, we develop a hybrid convolutional neural network with long short-term memory (CNN-LSTM), which effectively includes both spatial and temporal dimensions of time series data. In the second stage, a bidirectional real-time energy sharing strategy is designed based on forecasted data, facilitating the efficient distribution of surplus energy among communities. The two-stage system model integrates forecasting and energy sharing to accurately predict supply and demand, as well as effective energy sharing among GICN participants. The measurements including mean absolute error (MAE), root mean squared error (RMSE), and generation utilization rate are defined to evaluate prediction accuracy and energy sharing effectiveness thereby ensuring optimal grid operation and sustainability. Finally, the proposed hybrid CNN-LSTM algorithm is compared with the single LSTM and CNN models to demonstrate the performance superiority of the hybrid CNN-LSTM model. Numerical simulation experiments indicate that the proposed model CNN-LSTM achieved high accuracy in predicting user PV, wind, and load demand. Additionally, the implemented real-time energy sharing strategy efficiently manages energy distribution within GICNs.