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

01 Business Background

NIO, founded in November 2014, aims to build a global innovative smart energy service system. The focus of this article is the NIO Power business, which provides a network of charging and battery‑swap facilities powered by NIO Cloud technology, offering services such as home charging piles, fast chargers, swap stations, and one‑click charging via the NIO app.

02 Time‑Series Forecasting Background

Time‑series data are sequences that vary over time. Analysis can be performed in the time domain (temporal) or frequency domain. Typical patterns include periodicity, seasonality, and trend. Frequency‑domain analysis reveals dominant frequencies via spectral density.

03 Forecasting Tasks

Based on input variable count: univariate vs. multivariate.

Based on output length: single‑step vs. multi‑step.

Based on forecast horizon: short‑term, mid‑term, long‑term.

Typical application scenarios for NIO’s swap stations include new‑site selection, peak‑shaving charging, and battery dispatch, each requiring predictions at different horizons (24 h, 30 days, 12 months).

04 Key Challenges

Complex seasonal patterns across multiple stations.

Drifting time features such as non‑fixed holidays.

Growth and competition effects causing abrupt demand changes.

05 Algorithm Practice

System Architecture

Data are stored in a data warehouse and include attributes, operations, orders, users, vehicles, and weather. Feature engineering uses a feature engine (distribution, periodicity, related variables) and an embedding engine (token, value, positional, temporal embeddings).

Model Evolution

Statistical models: ARIMA, Prophet.

Machine‑learning models: LightGBM.

Deep‑learning models: TCN, CRNN, Informer, DCN.

Embedding details:

Token Embedding encodes station identifiers.

Value Embedding captures competition and growth variables.

Positional Embedding handles complex seasonality.

Temporal Embedding addresses holiday drift and time‑prior knowledge.

Model Fusion

Fusion strategies consider additive vs. subtractive methods, regression vs. classification, and reinforcement‑learning feedback loops.

Practical Results

Evaluation uses MAE and MAPE; after several iterations, MAPE stabilizes around 23 % (target < 5 %). Visual comparisons show LGB over‑fitting holidays, Informer struggling with long‑term seasonality, while DCN handles holiday alignment better.

06 Summary and Outlook

Future plans focus on faster real‑time updates, higher development efficiency via low‑code algorithm libraries, continued pursuit of algorithmic excellence, expanding functionality beyond forecasting, and open‑source collaboration.

07 Q&A

Q1: How large must a deep‑learning training set be to outperform traditional models?

A1: Approximately three years of historical data are used; deep models show clear advantages after stations have 2‑3 years of operation.

Q2: Will traffic‑flow prediction remain valuable after autonomous driving becomes widespread?

A2: Yes, forecasting peak periods and special events remains crucial for efficient vehicle‑network scheduling.