Air pollution, particularly PM2.5, is a major concern for public health in urban areas. In this paper, we propose a TinyML pipeline for PM2.5 forecasting that integrates a lightweight 1D-CNN with zero-phase Butterworth low-pass denoising to suppress sensor noise without phase distortion. Using hourly CPCB data from Perungudi (Chennai, 2024), the denoised CNN achieves strong accuracy on the normalized scale (MSE 0.000061, MAE 0.00389, RMSE 0.0078, R2=0.9908). The model is converted to TensorFlow Lite and deployed on an ARM Cortex-M4 with a 42 KB footprint, 11.4 ms inference latency, and 4,344 B RAM; measured power is 2,501 μ W. Relative to an unfiltered CNN and Multilayer Perceptron (MLP) / Feed-Forward Neural Network (FFNN) baselines, the denoising front end yields substantial error reductions while also lowering latency and memory. These results indicate that principled signal preprocessing paired with compact CNNs enables accurate, low-power, on-device P M2.5 monitoring suitable for smart-city deployments.