{"id":"39ee081c-c735-43e4-b7e1-c1c0cd27582e","arxiv_id":"1908.10713","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new unsupervised NILM algorithm, DUE, combines Markov-model activity chains with load simulation to disaggregate 15-minute smart-meter data into eight appliance categories, benchmarked on three European datasets.","lead":"The authors present an unsupervised algorithm that splits a household's 15-minute smart-meter reading into eight appliance categories using simulated daily activity chains, without labeled training data. It is a candidate tool for demand-side management, since it runs on common low-rate meters and needs only basic household information plus a time-use survey.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark uses sum-of-submeters, not real whole-house smart meter data; the claimed 15-min smart-meter performance is untested.","rationale":"The reader's weakest assumption was the generalizability of Dutch time-use statistics to central Europeans. That is a legitimate concern, but it is not the most load-bearing one. The reported experiments never actually feed DUE a real smart-meter signal: Section 2.2.2 explicitly sums the sub-metered appliance readings to form the 'whole-house' reference. Since DUE is designed and motivated for 15-minute smart meter data, the evaluation should use a real mains measurement. The paper itself acknowledges the potential impact of partial coverage in Section 3.3, yet does not perform such a test. Without that test, the 'under 20% uncertainty' result could simply reflect that the input is a perfect sum of the target categories, and therefore partially self-consistent. Even if Dutch TUS behavior is perfectly representative, the algorithm's performance on a constructed sum-of-parts signal does not establish its usefulness on real household mains. I therefore maintain the CONDITIONAL verdict, with the condition expanded to require validation on an actual whole-house signal. This is a different concern from the one identified by the reader, hence agreement_with_reader is 'disagree'.","tokens_in":30493,"tokens_out":5376,"duration_ms":55541,"concrete_test":"Run DUE on the actual whole-house active-power channel (mains) of UK-DALE House 3, or another dataset with a true mains reading, using the same metadata and appliance inventories, and compare per-category energy shares against sub-metered ground truth, adding an explicit 'unmonitored' residual if needed. If absolute Energy Share Errors remain below 20% for the categories present, the central claim survives; if they grow substantially, the current benchmark only demonstrates self-consistency on a constructed sum-of-parts signal.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 2.2.2 constructs the input signal as the sum of the monitored appliance sub-measurements: 'The reference whole-house power measurement is defined as the sum of the sub-measurements in order to ensure energy conservation.' All three datasets are down-sampled appliance-level recordings, so the 'whole-house' signal is a synthetic aggregate of the same category-level signals the algorithm is supposed to recover. Real 15-min smart meter data includes unmonitored loads, always-on devices, spikes, and noise, and Section 3.3 admits that partial coverage 'might have a significant impact on the performance of the algorithm.' Because this mismatch is not tested, the headline claim—accurate category shares from actual smart meter signals with uncertainties under 20%—is not supported by the reported experiments. The post-hoc 0.25 light-scaling factor (Section 2.1.3) further undermines out-of-sample claims, but the synthetic-aggregate benchmark is the more fundamental threat to the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised, hybrid load-simulation/disaggregation algorithm called device usage estimation (DUE). The method takes as inputs a 15-minute aggregate power measurement, household characteristics (number and age of inhabitants, employment state, appliance inventory, and usage habits), and activity statistics from the 2005 Dutch time-use survey. It builds Markov activity chains for each inhabitant, simulates appliance use category by category (Cooking, Entertainment, Fridge, Heating, Housekeeping, ICT, Light, Standby), and iteratively adjusts the simulated signal to match the measured aggregate. The output is the share of energy consumed by each category over the test period. The authors benchmark DUE against four NILM algorithms (CO, FHMM, DDSC, GSP) on three public datasets (ECO, SMARTENERGY.KOM, UK-DALE), downsampled to 15 minutes, and report energy share errors, estimation accuracy, and execution times. They conclude that DUE is unsupervised, has lower computational cost than most supervised benchmarks, and achieves prediction uncertainties that are normally under 20%.","tokens_in":30667,"tokens_out":8337,"duration_ms":74288,"significance":"If the claims are substantiated, this is a useful contribution to low-sampling-rate NILM. The person-centric activity-chain approach is novel, the target problem (category-level energy shares for demand-side management) is practically relevant, and the use of three public datasets with several baselines is a strength. The paper is clearly structured and provides enough algorithmic detail in the appendices to be implementable. The main value would be a genuinely unsupervised method that works at a 15-minute sampling rate, where most unsupervised NILM methods require higher-frequency data. However, the current experimental design does not fully establish the headline claim: the test signal is a synthetic sum of the very appliance channels the algorithm is supposed to recover, the algorithm is stochastic but evaluated with single runs, and several parameters are tuned on the test data. These gaps currently limit the strength of the significance claim.","major_comments":[{"comment":"The experimental setup does not test the claimed use case. Section 2.2.2 states: 'The reference whole-house power measurement is defined as the sum of the sub-measurements in order to ensure energy conservation.' All three datasets are downsampled appliance-level recordings, so the 'whole-house' signal is a synthetic aggregate of the same category-level signals that DUE is meant to recover. Real 15-minute smart meter data includes unmonitored loads, always-on devices, spikes, and noise; Section 3.3 acknowledges that partial coverage 'might have a significant impact on the performance of the algorithm.' Because this mismatch is not tested, the central claim that DUE disaggregates real smart meter signals with uncertainties normally under 20% is not supported by the reported experiments. A concrete fix is to run the same benchmark on the actual whole-house channels available in UK-DALE and ECO, or to add a controlled experiment with artificial unmonitored loads.","section":"Section 2.2.2 and Section 3.3"},{"comment":"The light-scaling factor and the fridge recalibration are post-hoc adjustments made after observing the algorithm's errors. The text in Section 2.1.3 states that the factor was added 'after noticing that our method over-estimated the light consumption,' and Appendix A, Eq. (35) recalculates the fridge nominal power from the night-time aggregate measurement (Pfridge_new = Pfridge_old * fridge_mean_power / (Pfridge_old * beta2)). These are free parameters fitted on the test data itself. Because the paper reports no sensitivity analysis and no out-of-sample validation of these choices, the reported uncertainties may be optimistic. The authors should either perform a cross-validation-style analysis of these parameters or clearly distinguish 'unsupervised' from 'parameter-free' and report the sensitivity of the results to these adjustments.","section":"Section 2.1.3 and Appendix A, Eq. (35)"},{"comment":"The abstract and conclusion state that prediction uncertainties 'normally remain under 20%,' but Table 5 shows several DUE energy share errors well above 20%: Entertainment errors are 33% on SMARTENERGY.KOM and 37% on UK-DALE, Light is 22% on ECO, and Housekeeping is -17% where measured. If 'uncertainty' refers to a different aggregate metric (for example, the average absolute error across categories in Figure 6), that definition should be given explicitly and the per-dataset values should be reported. As written, the claim is not supported by the data in the table.","section":"Section 3.1, Table 5"},{"comment":"The DUE algorithm is stochastic: activity chains are generated with random numbers (Appendix B), device durations are drawn as tau*2*epsilon, the device ordering is randomized (Section 2.1.2), and the activation variables gamma are random. Yet Tables 5 and 6 report single values for each metric without variance or confidence intervals. Without repeated runs, the reader cannot judge whether the reported performance is stable, and the 'uncertainty' of the method itself is not characterized. The authors should report means and standard deviations over at least 10-20 independent runs per household and dataset, and compare those distributions with the deterministic baselines.","section":"Section 2.1 and Appendix B"},{"comment":"The transferability of the Dutch time-use survey to central European households is acknowledged as questionable but is not validated. The sentence 'Although it is questionable if the behavior of Dutch citizens is representative of the behavior of all central Europeans, it seems like a reasonable assumption' is a single unsupported assertion. Since the activity chains drive the entire category allocation, an incorrect activity model would compromise the method's output. A concrete test is to repeat the experiments with the UK time-use survey (which the authors themselves cite in Section 2.1) or to compare the predicted activity/occupancy patterns with the actual occupancy information available in the ECO and SMARTENERGY.KOM datasets.","section":"Section 2.1"}],"minor_comments":[{"comment":"The phrase 'similar range of uncertainty while saving on the cost of acquiring labeled data' is stronger than the evidence: Table 5 shows several DUE errors above 20% and some negative Estimation Accuracy values. The claims should be tempered to match the reported results.","section":"Abstract and Conclusion"},{"comment":"The notation in the Kronecker delta conditions is confusing and may contain typos; for example, the condition 't0_i = 22:22' in Eq. (2) is likely meant to be a specific time, but it is not explained, and the use of backslashes instead of logical 'and' is nonstandard. The equations should be rewritten for clarity.","section":"Equations (2) and (3)"},{"comment":"The random number generator is not seeded and the algorithm is not accompanied by release code. For reproducibility, the authors should provide the code or, at minimum, describe the random seed policy and report variability across seeds.","section":"Section 2.1.2 and Appendix B"},{"comment":"The footnotes in Table 3 are difficult to parse; several footnotes are mixed in the 'Note' column and some references (e.g., 'footnote 3,' 'footnote 4') are ambiguous. The table would benefit from clear, separately numbered notes and a concise explanation of the beta and tau parameters.","section":"Table 3"},{"comment":"The legend in Figure 6 appears to contain a typo (it reads 'DUP' rather than 'DUE' in the provided text), and the y-axis is cropped without indication of which values fall outside the range. Please verify the legend and consider annotating clipped bars.","section":"Figure 6"},{"comment":"There are several typographical errors in the references (e.g., 'Methodies' instead of 'Methodologies' in [32]) and inconsistent page ranges. A careful proofreading pass is needed.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The synthetic-aggregate benchmark is the most fundamental threat to the paper's central claim, and I agree with the stress-test note on this point. The paper could be made publishable by re-running on real whole-house data (UK-DALE and ECO both have whole-house measurements), reporting the variance of the stochastic algorithm over multiple runs, and performing a sensitivity analysis for the light-scaling factor and fridge recalibration. The paper is from 2019 and the field has moved, but the method is still relevant if the claims are properly scoped. I would encourage the editor to request these revisions rather than reject outright, as the core idea is novel and the limitations are identifiable and potentially addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new unsupervised NILM approach with an honest benchmark, but the experiments use a synthetic whole-house signal built from the same appliance measurements the algorithm is supposed to recover, and a couple of parameters were tuned after seeing the results. That leaves the headline claim about real 15-minute smart meter data unsupported.\n\nThe novel piece is the combination: Markov activity chains from a time-use survey, household demographics, and iterative load-profile subtraction. I don't know of another method that does person-centric disaggregation at 15-minute sampling without labels. The authors also run a fair comparison against CO, FHMM, DDSC, and GSP on three public datasets, and they are upfront about the limitations of their survey data and dataset coverage. That alone is worth crediting.\n\nThe soft spots are real. First, and most important: Section 2.2.2 defines the 'whole-house' signal as the sum of the sub-measurements. That means the algorithm never sees unmonitored loads, always-on devices, or the noise a real smart meter captures. The paper admits partial coverage 'might have a significant impact' but nowhere tests how DUE behaves when the aggregate contains energy outside the eight categories. Until that is tested, the central claim is plausible but unproven.\n\nSecond, the 2.25 light-scaling factor was added after the authors noticed overestimation. That is post-hoc tuning on the test data, and it weakens the out-of-sample claim. The fridge nominal power is also recalibrated from the night-time signal of the same aggregate. These aren't disqualifying—every NILM method has some tuning—but they should be flagged as parameters, not model outputs.\n\nThird, the stochasticity. DUE uses random activity chains and random device ordering, but all results are single runs with no variance. Release of code and seeds, plus repeated-run statistics, should be a condition of acceptance. The conclusion says uncertainties 'normally remain under 20%,' but Table 5 shows DUE errors of 33% on Entertainment and -35% on ICT for one dataset. 'Normally' is doing a lot of work.\n\nAll that said, the core idea is worth pursuing. The paper deserves a serious refereeing, and I would ask for the reproducibility fixes and a real-whole-house test or a reasoned argument for why the sum-of-submeters benchmark is sufficient. The authors have built something new, and they seem aware of its limits.","headline":"Novel unsupervised NILM hybrid with an honest three-dataset benchmark, but the synthetic whole-house signal and post-hoc tuning leave the real-smart-meter claim unproven.","tokens_in":31206,"tokens_out":3215,"would_cite":true,"duration_ms":32882,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An unsupervised algorithm called device usage estimation splits a household's 15-minute power readings into the energy shares of eight appliance categories, with errors usually below 20 percent.","keywords":["non-intrusive load monitoring","energy disaggregation","unsupervised learning","Markov model","smart meters","demand-side management","low sampling rate","load profile simulation"],"falsifier":"Sub-meter every appliance in a set of households for at least four weeks, downsample to 15 minutes, run DUE with complete appliance inventories, and compare predicted category energy shares with the submetered truth; regularly exceeding 20 percent mean absolute energy-share error, or failing to match activity timing against resident diaries, would falsify the central claim.","tokens_in":30293,"feed_emoji":"⚡","tokens_out":16859,"duration_ms":140222,"temperature":0.7,"pith_summary":"The paper proposes device usage estimation (DUE), an unsupervised disaggregation method that takes a whole-house power signal sampled every 15 minutes plus household metadata—ages and employment state, heating type, appliance inventory—and returns the share of energy consumed by eight appliance categories over time. It builds a Markov-chain activity sequence for each adult inhabitant from the statistics of a time-use survey, then converts those activities into candidate appliance runs using nominal powers and usage probabilities, and subtracts them from the measured load until the energy budget is matched. The claim is that this hybrid of load simulation and disaggregation reaches category-level energy-share uncertainties normally under 20 percent, without needing any labeled data and with computational cost lower than most supervised benchmarks. The purpose is demand-side management: a utility could estimate the flexibility of many households and give tailored advice from ordinary smart-meter readings and a short survey, instead of installing sub-metering equipment.","feed_headline":"No-training method splits household power use into 8 categories","feed_subtitle":"No labels needed: 15-minute meter data plus household metadata, errors usually under 20 percent.","key_machinery":"The carrying mechanism is the activity chain, a Markov chain over 34 activity states whose transition matrix and initial distribution depend on the type of day, the employment state, and the age group of the inhabitant; these probabilities come from the Netherlands time-of-use survey. For each person over 10, the algorithm samples a random chain, maps activity episodes to appliance runs through a household appliance inventory with nominal powers and usage probabilities, and repeatedly re-samples until the simulated load matches the measured load within a tolerance. The category recognition is sequential—heating, lighting, cooking, housekeeping, entertainment, ICT—after standby and fridge are filtered out as the day's minimum and the periodic night-time refrigeration pattern respectively.","core_discovery":"The central claim is that disaggregating a low-sampling-rate household load into appliance categories does not require supervised training on labeled appliance data. A generative Markov model of human activity, parameterized by a generic time-use survey and by each household's composition, plus a simulation of appliance power draws, can be fitted to the measured whole-house curve well enough that the energy share per category is typically within 20 percent of the ground truth. The paper demonstrates this on three public datasets down-sampled to 15-minute resolution and benchmarks it against four standard NILM algorithms; while the supervised algorithms score higher on average, the unsupervised DUE stays in a similar uncertainty range and needs no training data.","pith_inferences":["The paper's own Dutch-survey caveat implies a direct empirical test: re-estimate the Markov activity statistics from a central European time-use survey and check whether the category-share errors shrink; the paper does not run this test.","The current all-inhabitant occupancy rule (any peak means everyone is home) is a likely error source; using occupancy sensors or per-person diaries would probably tighten the estimates.","Because the benchmark datasets monitor only a subset of household appliances, some of the reported error likely reflects the mismatch between the algorithm's assumed full appliance inventory and the dataset's partial coverage, rather than the method's intrinsic accuracy."],"forward_implications":["A utility could offer appliance-category-level feedback and demand-side management advice using only existing 15-minute smart meter readings plus a short household survey, without sub-metering.","The approach transfers to a new country or region by substituting a local time-use survey, since the activity statistics are the only behavior-specific input.","On the tested datasets the category energy-share uncertainty remains normally under 20 percent, the same range as the supervised benchmarks despite using no labels.","Execution time grows linearly with the length of the dataset, so long monitoring periods and large household pools are affordable compared with methods whose cost grows faster."],"supporting_citations":[{"why":"Defines the non-intrusive load monitoring problem and supplies the combinatorial optimization baseline used in the benchmark.","marker":"[2]"},{"why":"Supplies the discriminative sparse coding (DDSC) baseline, a supervised low-frequency disaggregation method compared against DUE.","marker":"[18]"},{"why":"Supplies the factorial hidden Markov model baseline, the main unsupervised comparison point at low sampling rate.","marker":"[24]"},{"why":"Provides the toolkit used to implement and run the combinatorial optimization and FHMM benchmarks on the test datasets.","marker":"[25]"},{"why":"Supplies the graph signal processing baseline for disaggregation at low sampling rates.","marker":"[28]"},{"why":"Source of the time-use statistics from which the Markov activity chains and duration distributions are estimated.","marker":"[32]"},{"why":"First test dataset, providing appliance-level ground truth used to compute category energy shares and accuracy.","marker":"[43]"},{"why":"Second test dataset, providing appliance-level ground truth from apartment deployments.","marker":"[44]"},{"why":"Third test dataset, providing whole-house and appliance-level readings for a longer evaluation period.","marker":"[45]"}],"fun_headline_variants":["Unsupervised NILM: home power mix from 15-min smart meter data","Label-free home energy disaggregation from low-rate meter data","No labels, just data: algorithm reveals household energy split","Household power breakdown without training data","Smart meter data alone yields appliance categories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Dutch citizens' daily activity statistics transfer to central European households; if that is false, the generated activity chains that drive the entire category allocation are disconnected from the households being disaggregated.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised NILM: home power mix from 15-min smart meter data","Label-free home energy disaggregation from low-rate meter data","No labels, just data: algorithm reveals household energy split","Household power breakdown without training data","Smart meter data alone yields appliance categories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3474,"prompt_tokens":968,"completion_tokens":2506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2426}},"tokens_in":584,"tokens_out":2506,"duration_ms":19710,"temperature":1.0,"reasoning_tokens":2426,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:18:55.844585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sub-meter every appliance in a set of households for at least four weeks, downsample to 15 minutes, run DUE with complete appliance inventories, and compare predicted category energy shares with the submetered truth; regularly exceeding 20 percent mean absolute energy-share error, or failing to match activity timing against resident diaries, would falsify the central claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the non-intrusive load monitoring problem and supplies the combinatorial optimization baseline used in the benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the discriminative sparse coding (DDSC) baseline, a supervised low-frequency disaggregation method compared against DUE."},{"cited_title":"Kim , author M","cited_arxiv_id":null,"evidence_quote":"Supplies the factorial hidden Markov model baseline, the main unsupervised comparison point at low sampling rate."},{"cited_title":"Batra , author J","cited_arxiv_id":null,"evidence_quote":"Provides the toolkit used to implement and run the combinatorial optimization and FHMM benchmarks on the test datasets."},{"cited_title":"He , author L","cited_arxiv_id":null,"evidence_quote":"Supplies the graph signal processing baseline for disaggregation at low sampling rates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the time-use statistics from which the Markov activity chains and duration distributions are estimated."},{"cited_title":"Beckel , author W","cited_arxiv_id":null,"evidence_quote":"First test dataset, providing appliance-level ground truth used to compute category energy shares and accuracy."},{"cited_title":"Alhamoud , author F","cited_arxiv_id":null,"evidence_quote":"Second test dataset, providing appliance-level ground truth from apartment deployments."},{"cited_title":"Kelly , author W","cited_arxiv_id":null,"evidence_quote":"Third test dataset, providing whole-house and appliance-level readings for a longer evaluation period."}],"review_version":1}