Non-Intrusive Load Monitoring (NILM) with very low-frequency data from smart meters in Switzerland

• Comprehensive evaluation of six NILM algorithms (FHMM, CO, Seq2Point, SGN, DAE, BiLSTM) for Swiss residential energy disaggregation. • Novel analysis of algorithm performance across multiple sampling frequencies (1-min, 5-min, 15-min) relevant to Swiss smart meter deployment. • Deep learning models (SGN, Seq2Point) outperform traditional benchmarks like FHMM and CO. • First cross-dataset validation study between REFIT and ECO datasets showing significant performance variations across different building datasets. • Systematic assessment of algorithm reliability for different appliance types (dynamic, steady-state, mixed behavior) at low sampling frequencies. In Switzerland, the residential sector accounts for approximately 34 % of total energy consumption. Understanding the energy consumption patterns of residential customers is crucial for developing effective energy efficiency strategies. Traditional methods for collecting energy consumption data rely on installing smart meters on individual appliances, which can be intrusive and expensive. Non-Intrusive Load Monitoring (NILM) offers a promising alternative by disaggregating the total energy consumption of a household into individual appliance-level energy consumption data. The frequency of data from smart meters can vary depending on the specific type of smart meter and the utility company’s data collection practices. In general, smart meters can collect data at a very high frequency, often in the range of minutes or even seconds. However, it is typical for mass-deployed smart meters to collect data at a frequency of 15 min, e.g. in Europe and Switzerland. This paper explores different NILM algorithms suited to very low frequency (1-min to 15-min sampling rate) metered data. Particularly the performance of the different algorithms is trained and tested on (i) same and (ii) different dataset using smart meter data at (i) 1 min, (ii) 5 min and (iii) 15 min frequency. The performance of NILM algorithms is highly dependent on the appliance, sample period, and dataset specificity. SGN emerges as the most versatile algorithm, excelling in capturing both dynamic and steady-state appliance behaviors across different sampling rates and validation scenarios.