Earthquake prediction or forecasting is an important component of disaster management, especially for countries located in seismically active regions, such as Japan. In this work, a new hybrid ensemble model named AGLCXG + L (ARIMA, GRU, LSTM, CNN and XGBoost as base models and Linear Regression as a meta-model) is proposed for earthquake magnitude forecasting in Japan. The proposed model is trained using a dataset covering the records from 1970 to 2024. The proposed methodology addresses the challenges of earthquake forecasting by combining time-series analysis with deep learning techniques. From the results, it is inferred that the proposed model achieved a better result with the R-squared value of 0.9777, MSE of 0.0069 and the MAPE of 1.288. These results demonstrated that the proposed hybrid ensemble approach for predictions of earthquake magnitude is effective for current and future seismic forecasting systems and more effective disaster management.