Preciseness in weather prediction is vital for the science of meteorology, as this precision is needed in planning and taking decisions effectively. This article deals with how, by applying the Convolutional Neural Networks model, it could bring about potential enhancements in weather prediction precision when contrasted between models built on personal experience and transferred learning models. We used three different CNN architectures: a CNN designed for this task, utilizing a pre-trained VGG16 architecture alongside a ResNet152V2 model to enhance predictive accuracy in the domain. The used models are trained on the dataset having Multiple Classes of Weather Images. Using advance Data augmentation strategies resulted in improvement in the ability of models to generalize, while mixed precision training made the training process more robust. These models were evaluated by some of the primary performance metrics: correctness, depletion, and error matrices. Results: The results showcase a certain modified CNN functions efficiently and helps save computation. However, models like VGG16 and ResNet152V2 which are based on the transfer learning act as dependable baselines with a high precision rate. This article highlights the power of Convolutional Neural Networks (CNNs) in the range of weather forecasting, as well as the connection between customized model building and transfer learning in creating dependable prediction models. The results provides a solid proof of using deep learning in metrology, therefore fostering the opportunities for further study and application of these models in real-world meteorological forecasting scenarios.