FeederBW Releases Two Years of German Feeder Data

Low-voltage grids are where much of the energy transition becomes operationally hard: photovoltaic systems, heat pumps, batteries, storage heaters, and EV chargers all change local load and feed-in patterns, but public data at feeder level remains scarce. FeederBW addresses that gap with real measurements from Netze BW’s German distribution network, covering 200 low-voltage feeders from April 1, 2023 to March 31, 2025. The paper is not proposing a new forecasting model. Its contribution is a measurement dataset with enough temporal resolution, metadata, and validation detail to support model development and grid analysis.

Core Contribution

The main advance is the combination of three data streams that are usually separated or unavailable: one-minute electrical measurements at LV feeders, feeder metadata, and weather data. The resulting dataset contains active and reactive power, voltage, current, and power-factor information, plus feeder-level attributes such as housing units, installed photovoltaic capacity, EV chargers, heat pumps, storage heaters, public lighting, batteries, and industrial or commercial categories. The authors also include changes in low-carbon technology installations over time, which matters because a feeder’s measured behavior can shift when new PV or charging capacity is connected.

Figure 1 summarizes the data assembly process: feeder measurements, metadata, and weather data are collected into a data platform, filtered, and split into published records.

Technical Approach

The dataset is organized as compressed Apache Parquet files and CSV metadata. The measurement data is split into 200 files, one per feeder, with 40 Parquet files per compressed archive. The paper recommends reproducible train-test protocols: for location-based evaluation, feeders F_1–F_160 are training feeders and F_161–F_200 are test feeders, with a five-fold variant using blocks of 40 feeders. For temporal evaluation, the recommended split uses April 1, 2023 to March 31, 2024 for training and April 1, 2024 to March 31, 2025 for testing.

The visual analysis in the paper shows why the metadata is not just a convenience field. Weekly quantile profiles distinguish household-dominated feeders from feeders with storage heaters, PV systems, and industry or commerce. In the examples, feeder F_37 has 102 housing units and no other feeder metadata; feeder F_160 has 78 kW of installed storage heaters visible in high nighttime quantiles; feeder F_97 has 62–80 kW of PV over the two-year period; and feeder F_33 combines industry and commerce with 76.8 kW of PV. Longitudinal plots also show seasonal structure, including higher winter loads and higher summer feed-in, and one feeder with a 59.16 kW PV installation event on January 10, 2024.

Results and Analysis

The scale is the central result. The authors report an expected 210,528,000 one-minute measurements across 200 feeders over two years, with 209,189,649 rows actually present. That corresponds to 99.36% completeness and 0.64% missing rows. Missing values inside existing rows are much rarer for voltage, with 833 missing voltage measurements, reported as 0.0004%, but more common for PEN values at 25.6%. For 51 feeders, PEN values are absent, and F_181 has only about two-thirds of expected rows after postprocessing.

The dataset also captures visible adoption of low-carbon technologies. Figure 6 shows average installed power per LV feeder from April 1, 2023 to March 31, 2025: batteries increase by 148.2%, PV systems by 52.2%, EV chargers by 28.4%, heat pumps by 21.8%, hot-water tanks by 10.1%, flow-type heaters by 1.9%, storage heaters by 1.7%, and electric heaters by 0.4%. Public lighting, other consumers, and CHP remain at 0.0% change in this view.

For researchers, the useful feature is not only size but interpretability. A one-minute feeder series without metadata can support forecasting benchmarks, but it is weaker for studying why demand and feed-in change. FeederBW makes those shifts easier to connect to PV, heating, storage, EV infrastructure, and weather. That makes it relevant for load forecasting, synthetic data generation, non-intrusive load monitoring, and studies of low-voltage grid behavior under electrification.

Caveats in Practice

The paper is careful about measurement limits. Power values are approximate in low-current and harmonic scenarios because current accuracy is 3% and voltage is measured on one phase, which can affect derived phase power factors. Active, reactive, and apparent power do not always satisfy $S = \sqrt{P^2 + Q^2}$, partly because of opposite current and power flows across phases. The authors give concrete deviations: before filtering, the largest deviation is 38.06 kVA for feeder F_127 on April 9, 2023; after filtering by total power factor, the highest deviation is 9.46 kVA for F_9 on November 3, 2023.

Metadata also has constraints. Equipment in LV grids can be unregistered, registration dates may differ from operational start dates, and switching events can alter topology in ways that create undocumented concept drift. Weather records contain no missing rows, but snow-height nulls appear mainly from June to October 2024 and can usually be treated as 0 m. These caveats do not undermine the dataset; they define the conditions under which it should be used. FeederBW is strongest as a real-world benchmark for feeder-level temporal modeling, less suited to exact physical inference about cable capacity, topology, or phase-level power flows.