This thesis addresses the critical issue of underestimated wake effects between neighboring windparks by developing efficient long-distance wind farm flow models using Convolutional NeuralNetworks (CNNs). The study compares three wake deficit models (Jensen, Bastankhah, andTurbOPark) and four neural network architectures (Convolutional Autoencoder (CAE), U-Net,CAE/MLP, and U-Net/MLP) to improve long-distance wake predictions.A novel method for random wind park layout generation was developed, simulating diversescenarios up to 768 rotor diameters downstream. Each wake model dataset comprised 1000simulations, split 80/20 for training and testing. Results demonstrate that all neural net-works effectively simulate wake datasets, with U-Net and U-Net/MLP consistently outperform-ing CAE approaches. Mean Absolute Errors (MAE) range from 4.75 × 10−4 m/s (Jensen) to1.44 × 10−2 m/s (TurbOPark). The U-Net/MLP model also successfully predicted turbulenceintensity, achieving MAEs between 1.30 × 10−4 (Frandsen model) and 2.21 × 10−4 (Crespo-Hernández model).Crucially, neural networks significantly outperform traditional engineering models in compu-tational efficiency. While engineering models’ computational time scales linearly with turbinecount, neural networks maintain a constant execution time of approximately 3 ms, regardlessof wind park size. This breakthrough enables rapid assessment of large-scale wind farm layoutsand performance optimization.