Benchmarking neural architectures for long-horizon forecasting in ecological simulations

Ecological and biological systems exhibit complex nonlinear temporal dynamics that challenge traditional time series forecasting approaches. In this work, we compare multilayer perceptrons (MLPs), convolutional neural networks (CNNs), and graph-based neural architectures for predicting long-term population trajectories in a simulated avian ecosystem. We focus on (i) convolutional temporal filters for capturing seasonal and local dependencies, (ii) assessing whether graph representations of spatial structure improve forecasting performance, and (iii) quantifying robustness under monitoring constraints via data ablations (reduced runs, weekly sampling, reduced scenario diversity). We evaluate model accuracy across multiple forecasting horizons and analyse performance differences across architectures on a large-scale simulation dataset.