Forecasting intraday prices in commodity markets, such as the oil market, is technically challenging due to their high volatility, nonlinearity, and sensitivity to external events. Traditional econometric models, such as ARIMA, have been surpassed by deep learning (DL) techniques, including long short-term memory (LSTM) networks, convolutional neural networks (CNNs), and multi-layer perceptrons (MLPs), in terms of predictive capacity, particularly with regard to intraday prices. However, these models are often limited in terms of interpretability and computational efficiency. This study compares the performance of three DL architectures-MLP, LSTM, and KAN-for intraday oil price forecasting. We evaluate metrics such as accuracy, stability, and computational cost to identify the strengths of each approach in short-term scenarios. Empirical results demonstrate that although MLP and LSTM networks fit the data well, KANs offer additional advantages, including greater accuracy and the capacity to capture abrupt changes without requiring complex architectures. Their compositional capacity and hierarchical structure enable them to efficiently represent nonlinear dynamics, a useful capability in highly variable energy markets.