I need a complete deep-learning workflow that predicts future values for a multichannel time-series dataset. The raw data are already cleaned and timestamped; what I don’t have is the modelling logic, training script, and evaluation framework that will let me hand off a reliable forecast to my supervisor. My main objective is straightforward: accurate forecasting. You are welcome to engineer features, experiment with different sequence lengths, or fold exogenous variables into the network, but the final solution must be a deep-learning model (e.g., LSTM, GRU, Temporal CNN, Transformer). Classic statistical or shallow machine-learning baselines may be used only as references to justify the deep-learning gains. Because the data domain itself is not the bottleneck, please structure the code so the input pipeline can accept any comma-delimited file with a time column plus N numeric channels. A clean README that explains hyper-parameters, training requirements, and how to swap in a new dataset will save me a lot of time when I demonstrate the system internally. Deliverables • Well-commented Python code (TensorFlow or PyTorch) that trains, validates, and tests the model • Saved model weights ready for inference • A short report (Markdown or PDF) summarising architecture choices, key metrics, and suggested next steps • Optional: a notebook with visualisations that compare forecasted versus actual values on the hold-out set Acceptance criteria • Reproducible results on my machine (Ubuntu 22.04 / Python 3.10) • Clear performance metrics such as MAE or RMSE logged per epoch and final • Inference script returns a forecast array for an arbitrary horizon without manual tweaks If you’ve built similar forecasting pipelines—or have creative ideas like attention-based encoders—let me know in your first message along with a link to a relevant repo or demo.