This paper focuses on assessing the capability of Kolmogorov-Arnold Networks (KAN), a new type of neural network introduced by MIT, in comparison with Multilayer Perceptrons (MLPs) and other traditional deep learning models for time series forecasting, specifically for predicting rainfall in the state of Telangana. KAN is based on the Kolmogorov-Arnold Representation Theorem and consists of learnable activation functions, unlike traditional deep learning models, which eliminates the need for weights and biases. We aim to implement KAN and Temporal KAN (TKAN), a version of KAN designed to handle temporal dependencies, along with Neural Networks, MLPs, Convolutional Neural Networks (CNNs), and Long Short-Term Memory Networks (LSTMs) to compare their forecasting accuracy, parameter efficiency, and training time. The dataset used consisted of nearly 24 months of rainfall data from three districts in Telangana, namely Rangareddy, Sangareddy, and Hyderabad. Our results show that TKAN outperformed all models, with KAN ranking second in most cases, indicating their potential for greater forecasting accuracy and advancing the field of predictive modeling. Additionally, KANs were found to be more parameter-efficient than all models, except for neural networks. The ability of KANs to model complex nonlinear relationships and temporal dependencies, while offering greater parameter efficiency, highlights their potential to deliver superior accuracy and interpretability, making them a promising alternative to orthodox models like MLPs.