Foreign exchange rates play a significant role in global finance, impacting international trade, investment decisions, and economic stability. Due to their volatile and nonlinear behavior, accurately predicting currency exchange rates has become a crucial area of financial research. This study analyzes historical exchange rate data to identify complex temporal and spatial patterns using advanced deep learning techniques. In this study, the main objective is to evaluate and compare the predictive performance of five deep learning models for EUR/USD exchange rate prediction. The model architectures are Convolutional Neural Network (CNN), Multi-Kernel Convolutional Neural Network (Multi-CNN), Attention Mechanism (AM), and two hybrid models, namely, Deep Attentive Convolutional Fusion (DACF) and Deep Attentive Multi-Kernel Convolutional Fusion (DAMCF). A big data analysis is conducted using 5,352 daily Open, High, Low, and Close (OHLC) values for the EUR/USD exchange rate covering the period from 2003 to 2024. Weighted averaging and normalization were used as preprocessing techniques. Then, the methodology involves rolling window analysis, candlestick chart visualization, and the design of model architectures that include CNN layers, multi-scale convolution kernels, attention mechanisms, and hybrid models combining spatial and temporal features. The models are evaluated using RMSE, MAPE, and R². The experimental results show that all the models achieved considerable accuracy, with the three best performing models being DAMCF (R² = 0.9556), DACF (R² = 0.9210), and AM (R²=0.9051). Further, the DAMCF model demonstrates the greatest ability to learn complex market patterns. These results demonstrate the potential of hybrid architectures in financial forecasting that combine CNN and AM. The suggested method provides insightful information for monetary policy planning, risk management, and algorithmic trading.