ZJU-o and Downstream Benchmark Datasets: Battery Operational Time Series and Semantic Reports for LLM-Assisted O&M

# ZJU BESS Time-Series, Semantic Reports, and Downstream Benchmark Suite ## Overview This dataset suite provides large-scale, real-world operational time series from a lithium-ion battery energy storage system (BESS), together with structured, human-readable operational reports and three ready-to-use downstream benchmarks. Unlike battery datasets designed for a single prediction target, this release connects three levels of battery intelligence: 1. **Raw multivariate operation data** from 28 battery modules; 2. **Structured semantic reports** describing module- and cell-level temporal behavior; 3. **Task-specific labels** for SOC prediction, operational-anomaly analysis, and charge/discharge decision support. The suite is intended to support research at the intersection of battery management, multivariate time-series analysis, explainable AI, and large language models (LLMs). It enables both conventional numerical modeling and emerging time-series-to-language research without requiring users to reconstruct benchmark samples from the full operational archive. The operational records cover a six-month period from **December 1, 2023 to May 31, 2024**. Measurements are sampled at one-minute intervals. All public identifiers are anonymized and independently assigned within the released datasets. ## Dataset Components | Dataset | Purpose | Samples | Data and labels | |---|---:|---:|---| | **ZJU-o** | Large-scale operational time-series and semantic-report corpus | 72,520 windows | Multivariate battery time series paired with module- and cell-level reports | | **ZJU-t** | SOC prediction benchmark | 100 samples | Time series with module-average SOC prediction targets | | **ZJU-ta** | Operational-anomaly analysis benchmark | 130 samples | 100 normal and 30 anomalous samples with anomaly-category labels | | **ZJU-td** | Charge/discharge decision-support benchmark | 40 samples | Historical time series with time-to-cutoff labels | ### ZJU-o: large-scale operational and semantic-report corpus ZJU-o contains **72,520 operational windows** from 28 anonymized battery modules. Each module contributes 2,590 windows. Every window covers 100 consecutive timestamps and is paired with five structured reports generated through the TS2R reporting pipeline: - one module-level report describing the overall operating behavior; - four cell-level reports covering cells 1--4, 5--8, 9--12, and 13--16. The reports organize quantitative battery behavior into readable descriptions of operating trends, transitions, fluctuations, outliers, inconsistencies, and key statistical values. In total, ZJU-o provides **362,600 aligned operational reports**, making it suitable for large-scale time-series-to-text and battery-language research. ### ZJU-t: SOC prediction benchmark ZJU-t contains 100 labeled samples for SOC prediction. Each sample provides a complete 100-minute multivariate time-series window. The label is the **mean SOC across all 16 cells at the 100th timestamp**. A standard forecasting protocol is to use timestamps 1--90 as observations and predict the label at timestamp 100. Retaining the complete window also allows users to define alternative observation lengths, conduct retrospective analysis, or design representation-learning tasks. ### ZJU-ta: operational-anomaly analysis benchmark ZJU-ta contains 130 labeled windows: - 100 `normal` samples; - 10 `voltage_cell_anomaly` samples; - 10 `temperature_cell_anomaly` samples; - 5 `soc_cell_anomaly` samples; - 5 `soc_inconsistency` samples. These labels describe **observable measurement or operational anomalies**. They should not be interpreted as confirmed physical root causes such as internal short circuits or thermal runaway. The benchmark supports binary anomaly detection, imbalanced classification, anomaly-category recognition, interpretable anomaly assessment, and comparison between numerical and language-based methods. ### ZJU-td: charge/discharge decision-support benchmark ZJU-td contains 40 samples for time-to-event forecasting during active charging or discharging. Each sample includes a 100-minute historical operating window. Its label is the number of minutes from the observation endpoint to the first time the predefined SOC cutoff condition is reached. The benchmark covers both charging and discharging trajectories, with cutoff conditions of approximately 97% SOC for charging and 3% SOC for discharging. It can be used to evaluate time-to-cutoff forecasting, early/late prediction behavior, operational scheduling support, and interpretable supervisory decision support. ## Common Time-Series Layout Every sample in the four datasets uses the same multivariate layout: - **Shape:** `49 x 100` (`channels x timestamps`); - **Sampling interval:** one minute; - **Cells:** 16 series-connected cells; - **Cell variables:** temperature, voltage, and State of Charge (SOC); - **Module variable:** one shared current signal. The 49 channels are ordered as: ```text Cell 1 temperature, Cell 1 voltage, Cell 1 SOC, Cell 2 temperature, Cell 2 voltage, Cell 2 SOC, ..., Cell 16 temperature, Cell 16 voltage, Cell 16 SOC, Module current ``` Equivalently, rows `0:48` contain 16 repeated groups of `[temperature, voltage, SOC]`, and row `48` contains module current. ## File Organization ### ZJU-o JSON files ZJU-o is organized into 28 anonymized module files: ```text data_for_module#1.json data_for_module#2.json ... data_for_module#28.json ``` Each file contains a JSON list of 2,590 records. Every record has the following fields: - `sample id`: sample identifier within the public module file; - `time series`: the `49 x 100` multivariate time-series array; - `descript c1`: module-level operational report; - `descript c2`: cell-level report for cells 1--4; - `descript c3`: cell-level report for cells 5--8; - `descript c4`: cell-level report for cells 9--12; - `descript c5`: cell-level report for cells 13--16. ### Downstream benchmark pickle files The three downstream benchmarks are distributed as: ```text ZJU-t.pkl ZJU-ta.pkl ZJU-td.pkl ``` Each pickle file contains a dictionary with a `samples` list. Every public sample contains only three fields: ```python { "id": 1, # sequential public identifier "data": array, # NumPy array with shape (49, 100) "label": value # task-specific numerical or categorical label } ``` The meaning of `label` is dataset-specific: - **ZJU-t:** module-average SOC at timestamp 100; - **ZJU-ta:** operational category (`normal` or one of four anomaly categories); - **ZJU-td:** remaining minutes to the SOC cutoff condition. ## Potential Uses This suite can support substantially more than report generation. Potential applications include: - multivariate battery time-series forecasting; - SOC estimation and forecasting; - operational-anomaly detection and anomaly-category recognition; - imbalanced and few-shot battery classification; - time-to-cutoff and time-to-event prediction; - charge/discharge scheduling and supervisory decision support; - temporal segmentation, change-point detection, and operating-state recognition; - self-supervised and contrastive representation learning for battery signals; - cross-module transfer learning and domain adaptation; - explainable battery management and evidence-grounded prediction; - time-series-to-text generation and automated engineering reporting; - time-series question answering and retrieval-augmented battery analysis; - multimodal alignment among numerical signals, plots, and natural language; - instruction tuning and evaluation of battery-oriented foundation models; - benchmarking numerical, visual, textual, and report-based model interfaces; - predictive maintenance research and human--AI engineering communication. The combination of a large operational corpus and compact labeled benchmarks allows researchers to pretrain or study representations on ZJU-o and evaluate downstream transfer on ZJU-t, ZJU-ta, and ZJU-td. ## Loading Examples ### Load one ZJU-o module ```python import json import numpy as np with open("data_for_module#1.json", "r", encoding="utf-8") as f: records = json.load(f) sample = records[0] time_series = np.asarray(sample["time series"], dtype=float) module_report = sample["descript c1"] print(time_series.shape) # (49, 100) print(module_report) ``` ### Load a downstream benchmark ```python import pickle with open("ZJU-ta.pkl", "rb") as f: dataset = pickle.load(f) samples = dataset["samples"] data = samples[0]["data"] label = samples[0]["label"] print(len(samples)) # 130 for ZJU-ta print(data.shape) # (49, 100) print(label) ``` ## Notes on Responsible Use - The data represent measurements from a specific real-world BESS and should not be assumed to cover every battery chemistry, system design, climate, or failure mechanism. - ZJU-ta provides operational-anomaly labels rather than verified physical fault-cause labels. - ZJU-td is intended for forecasting and supervisory decision-support research. It is not a replacement for certified BMS protection logic or hard safety constraints. - Pickle files should only be loaded from trusted sources. The files in this official release contain standard Python dictionaries, NumPy arrays, identifiers, and labels. ## Related Method The semantic reports were constructed using **TimeSeries2Report (TS2R)**, a signal-to-semantics-to-report pipeline that translates multivariate battery time series into structured operational descriptions. TS2R reports can be used directly as human-readable records or as semantic inputs to off-the-shelf LLMs for downstream battery operation and maintenance tasks. ## Citation and Attribution To acknowledge the efforts involved in data collection, curation, and annotation, and to ensure proper attribution and reproducibility, **any research using this dataset or any part of it MUST cite our associated TS2R paper**. **Associated Paper**: Yang, J., Guay, M., Cao, Z., and Zhao, C., A training-free semantic cognition framework unlocks interpretable, actionable management for battery energy storage systems, Cell Reports Physical Science, 2026.