The volatility of agricultural commodity prices poses a significant challenge to farmers and policymakers in Tamil Nadu, India. Accurate price forecasting is crucial for market stability and economic planning. This study proposes a novel deep learning ensemble model for predicting daily wholesale crop prices at a district-level. Price data for numerous commodities was sourced from the Agmarknet portal (2015-2025) and augmented with district-specific weather data (mean temperature, precipitation sum) from the Open-Meteo API. A key innovation of this work is the integration of weather data using variable-length lookback periods tailored to the agronomic timelines of each crop. We first establish a baseline using a Seasonal AutoRegressive Integrated Moving Average with eXogenous factors (SARIMAX) model. Subsequently, we develop and evaluate four deep learning models: a general Gated Recurrent Unit (GRU), an optimized GRU, a general Long Short-Term Memory (LSTM), and an optimized LSTM. The final predictive model is an ensemble that averages the outputs of these four deep learning models. The results indicate that the deep learning ensemble significantly outperforms the SARIMAX baseline across most commodities. For instance, the ensemble model for Garlic achieved a Mean Absolute Percentage Error (MAPE) of approximately 2.51%, demonstrating high accuracy. The study confirms that while deep learning models provide superior performance, the inclusion of tailored weather data offers a discernible, albeit variable, improvement, underscoring the importance of feature engineering guided by domain knowledge.

Keywords: Crop Price Prediction, Deep Learning, Ensemble Model, Time Series Forecasting, Weather Integration