Time Series Forecasting for NIO Power Swap Stations: Business Background, Challenges, and Algorithm Practice

NIO aims to build a globally innovative smart energy service system, with the NIO Power business providing a comprehensive charging and battery‑swap solution that integrates home chargers, fast chargers, super chargers, swap stations, and one‑click charging via the NIO app.

Business Background : NIO, founded in 2014, operates five major business blocks (NIO House, NIO Life, NIO Power, NIO Service, NIO Certified) and offers a range of electric vehicle models. The focus of this article is the NIO Power segment and its nationwide network of over 1,000 swap stations.

Time‑Series Forecasting Background : Time‑series data are sequences that vary over time. Forecasting explores both temporal and frequency domains to capture periodicity, seasonality, and trend. Visualizations of temporal and frequency domains illustrate these concepts.

Demand‑Prediction Tasks : Forecasting tasks are classified by input variable count (univariate vs. multivariate), output length (single‑step vs. multi‑step), and horizon (short‑term, mid‑term, long‑term). Typical applications include new‑site selection, peak‑shaving charging, and battery dispatch.

Key Challenges :

Complex multi‑sequence seasonality: different swap stations exhibit distinct daily cycles and holiday patterns.

Time‑feature drift: holidays shift yearly and future time variables are known while other features are not.

Growth and competition: expanding user bases around stations and the impact of new or closed stations cause demand spikes.

Algorithm Tasks : The three business needs translate into three Seq‑to‑Seq tasks—multivariate‑multivariate short‑term, mid‑term, and long‑term predictions.

System Architecture :

Data: stored in a data warehouse, covering attributes, operations, orders, users, vehicles, weather, etc.

Feature Engine: extracts distributional, periodic, and relational features.

Embedding Engine: includes Token, Value, Positional, and Temporal embeddings to handle station identity, competition, seasonality, and holiday drift.

Models: progression from ARIMA, Prophet, LightGBM to deep‑learning models such as TCN, CRNN, Informer, and DCN.

Model Fusion: combines base models using additive (stacking) or residual (subtraction) methods, and considers classification vs. regression for ensemble decisions.

Service: deployed via an algorithm platform for online production.

Model Evaluation : MAE and MAPE are used; after several iterations MAPE stabilizes around 23%, indicating a performance ceiling. Visual comparisons show LGB over‑fitting holidays, Informer struggling with long‑term seasonality, while DCN handles holiday alignment better.

Future Plans :

Real‑timeization: increase the proportion of real‑time features to improve prediction freshness.

Efficiency: create a low‑code algorithm library based on the unified architecture.

Value: extend the framework to cover more time‑series tasks, open‑source components, and empower other departments.

The article concludes with a Q&A discussing dataset scale for deep learning and the relevance of traffic flow prediction in an autonomous‑driving era.