**6. Conclusions**

The LIT-Dataset was presented chronologically. Its rationale in supporting our own NILM development as well as making it publicly available. Its conception; its requirements elicitation and specification, based on an evaluation of available NILM datasets and the additional needs. Its design, structuring the LIT-Dataset into three subsets, each exploring a different load-set context. The design and implementation of the supporting systems for each of the subsets: jig, simulator, and NSAS. Its evaluation and validation, based on the comparison of simulated loads to real-world loads as well as its usage in NILM techniques. Finally, its publication (the LIT-Dataset is publicly available, upon free registration, at http://dainf.ct.utfpr.edu.br/~douglas/LIT\_Dataset), with detailed documentation and usage scripts.

The three subsets consider the scenarios of (1) a set of up to eight loads that are controlled (on and off switching) individually during the recording of aggregated current and load events; (2) a set of simulated loads that are recorded under conditions that would be difficult in real-world situations, either because they are uncommon or due to hazardous scenarios such as short-circuits; and (3) a set of loads monitored during their daily use. The first subset is the named Synthetic load shaping, as the "on" and "off" events are controlled, the second is named Simulated, and the third is named Natural load shaping as there is no influence on the loads during the recording period.

Among the distinct features of the LIT-Dataset, as described in Section 5.4, is the labeling of the load events at sample level resolution and with an accuracy better than 5 ms; the availability of such precise timing information that also includes the identification of the load and of the sort of power event is an essential requirement both for the evaluation of NILM algorithms and techniques, as well as, for training of NILM systems, particularly those based on Machine Learning.

Our contribution is to make publicly available a new dataset whose combination of features makes it unique. These features are: (1) the availability of load-event labels, with an accuracy better than 5 ms, providing ground-truth information of the load events, (2) the availability of three subsets (as described above), (3) recording scenarios with up to eight concurrent loads, (4) combination of residential, commercial and low-voltage industrial loads, and (5) load shaping scenarios with low-power loads being switched when high-power loads are energized.

To summarize the benefits of these contributions, concerning the availability of load-event labels, the LIT-Dataset achieved the best accuracy among the datasets that were analyzed (Table 17). This is an important characteristic for those using a dataset to validate event detection and load classification algorithms. Having loads recorded individually and concurrently also provides the required information for training as well as for evaluating the performance of NILM algorithms. Furthermore, scenarios where low-power loads switching when higher-power loads are powered-on, provides challenging test cases for these NILM algorithms.

The LIT-Dataset was presented here, from its conception to implementation, analysis of results, and publication. However, data collection is in progress as new loads, and new scenarios are frequently recorded and added to the dataset.

**Author Contributions:** Conceptualization, D.P.B.R., F.P., H.C.A., and C.R.E.L.; data curation, D.P.B.R., F.P., H.C.A., R.R.L., L.d.S.N., L.T.L., B.M.M., and J.R.L.d.S.; formal analysis, D.P.B.R., A.E.L., R.R.L., and E.O.; funding acquisition, D.P.B.R., A.E.L., J.S.O., and R.B.d.S.; investigation, D.P.B.R., F.P., H.C.A., A.E.L., E.O., and L.d.S.N.; methodology, D.P.B.R., F.P., H.C.A., A.E.L., C.R.E.L., and R.R.L.; project administration, D.P.B.R., J.S.O., and R.B.d.S.; resources, D.P.B.R., C.R.E.L., and R.R.L.; software, D.P.B.R., H.C.A., C.R.E.L., R.R.L., L.T.L., and B.M.M.; supervision, D.P.B.R., J.S.O., and R.B.d.S.; validation, D.P.B.R., F.P., H.C.A., A.E.L., C.R.E.L., E.O., L.d.S.N., and L.T.L.; visualization, D.P.B.R., L.d.S.N., and J.R.L.d.S.; writing—original draft preparation, D.P.B.R., F.P., H.C.A., A.E.L., C.R.E.L., R.R.L., and E.O.; writing—review and editing, D.P.B.R., F.P., and E.O. All authors have read and agreed to the published version of the manuscript.

**Funding:** This study was fully financed by Agência Nacional de Energia Elétrica (ANEEL) and Companhia Paranaense de Energia Elétrica (COPEL) under the research and development program (project PD2866-0464/2017).

**Acknowledgments:** The authors would like to thank COPEL and ANEEL for the support and promotion in the research project PD2866-0464/2017.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; the authors affiliated to the funder company had the role of manuscript revision and evaluating the request for publication.
