Accurate crop yield prediction is crucial for optimizing agricultural practices, enhancing economic planning, and ensuring food security. This study presents an in-depth analysis of crop yield forecasting in Bangladesh using deep learning models. Time-series data from 1981 to 2024 were collected from multiple sources, including NASA Power Data Access Viewer, the Bangladesh Meteorological Department, and the Bangladesh Bureau of Statistics (BBS), covering key crops such as rice, jute, wheat, and potatoes. The proposed custom Bi-LSTM model was used to predict crop yields based on various meteorological and statistical parameters. Alongside it, three other deep learning models—LSTM, GRU, and 1D-CNN—were also employed for comparison. Experimental results indicate that the proposed Bi-LSTM model outperformed other architectures, achieving the lowest error rates (MSE: 0.00198, RMSE: 0.04459, MAE: 0.0264) and the highest R2 value (0.93735), demonstrating its superior capability in capturing temporal dependencies. This research contributes to the advancement of precision agriculture by identifying an optimal deep learning framework for reliable crop yield forecasting in Bangladesh.