Existing global numerical weather prediction models typically provide large-scale, low-resolution forecasts, which are insufficient to meet the requirements of localized, fine-scale meteorological prediction. This study focuses on forecasting the 30-day average temperature in Fujian Province and proposes two deep learning approaches based on a multi-layer perceptron (MLP) and a convolutional neural network (CNN). Experimental results show that the independently trained CNN model with a 5×5 grid input achieves the best performance, reducing the mean absolute error (MAE) and mean squared error (MSE) by 65.3% and 86.7%, respectively, and increasing the correlation coefficient by 4.4%, compared to the baseline provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). These findings demonstrate that selecting appropriate grid scales and model architectures can substantially improve the accuracy of temperature forecasts in specific regions.