NON-INTRUSIVE LOAD MONITORING FOR SMART GRIDS - Dell
NON-INTRUSIVE LOAD MONITORING FOR SMART GRIDS William Schneider Sr. Data Scientist Dell EMC William.Schneider@dell.com Fernanda Campello de Souza Sr. Data Scientist Dell EMC Fernanda.Campellodesouza@emc.com Knowledge Sharing Article 2018 Dell Inc. or its subsidiaries.
Table of Contents Abstract . 3 Introduction . 3 Data description . 4 Data exploration: aggregate energy consumption . 7 Peak-period loads . 9 Demand forecasting . 12 ARMA Model . 13 Baseline Model. 14 Conclusion and future work . 15 References . 17 Disclaimer: The views, processes or methodologies published in this article are those of the authors. They do not necessarily reflect Dell EMC’s views, processes or methodologies. 2018 Dell EMC Proven Professional Knowledge Sharing 2
Abstract Electric power distribution systems have received increased attention in the past decade due to greater availability of IoT devices, including two-way smart meters and edge gateways. This increased detail in data provides utilities with deeoer visibility into the behavior of the grid to support challenges such as demand response, where the utility must predict immediate demand since energy storage is limited. Another active research effort in the delivery of electricity is non-intrusive load monitoring (NILM), in which aggregate electricity usage data is used to determine the array of devices present. These two efforts, one on a macro scale and another on a micro scale, could be merged into an approach which has greater effectiveness. In this paper, we use the REFIT dataset to evaluate the potential of NILM techniques to support demand response efforts through time-of-use pricing and macro-forecasting. Additionally, we utilize unsupervised techniques to make the technique available to a much wider customer base. Introduction The term Smart Grid refers to electric power distribution systems equipped with sensors along transmission lines that can give real-time information on operation conditions and also enable two-way communication between the utilities and customers. The increased availability of detailed operation information in the smart grid allows for optimizing the process of energy generation and distribution, improving reliability and promoting efficient use of the current infrastructure to meet a growing demand for energy. More than half of utilities in the US are currently deploying smart grid infrastructure to support an array of business challenges. The sensor devices market, in particular, is a growing one that is expected to grow from USD 12.8 billion in 2017 to 20 billion by 2022 globally. Since, in most situations, electricity cannot be stored in the amounts necessary to meet typical demand gaps, its generation and distribution need to work as a just-in-time process, with more electricity generated when demand increases. This creates challenges for utilities at times of peak demand, sometimes forcing the use of more expensive and/or more polluting electricity generation methods, or purchasing electricity from neighboring grids at a premium. In worst-case scenarios, it may result in rolling or total blackouts. A potential benefit of smart grids is the capability to promote individual behavioral changes in power consumption that can smooth overall energy demand, avoiding surges that can increase costs and decrease reliability for all of a utility’s customers. This can be done through increasing a customer’s visibility on their energy consumption at different times and offering incentives for demand smoothing, such as time-of-use pricing programs. The main device that enables this detailed visibility on electricity consumption is the smart meter, an electricity meter that records aggregate consumption in short intervals (less than 1 hour) and transmits this data to the utility, allowing for detailed monitoring of seasonality in consumption. In order to help customers and utilities identify specific opportunities for savings, however, direct visibility into consumption of specific devices would be useful. Smart meters can be deployed as a single meter for a customer or in complex situations in order to measure specific devices. However, the field of NonIntrusive Load Monitoring (NILM) aims to disaggregate high-level aggregate measurements to contributions of individual appliances, based on a single metering point (smart meter at main breaker 2018 Dell EMC Proven Professional Knowledge Sharing 3
level). The latter promises to be a cheaper method of implementation and is therefore an active area of research. Non-Intrusive Load Monitoring has many published algorithms available, including both supervised methods that require expensive labeled (disaggregated) data for training, as well as unsupervised methods that, despite being generally less accurate than supervised methods, are of high interest due to low setup cost and short training phase. In addition, NILM techniques differ in their data requirements: energy, voltage, as well as frequency measurements. The methods most likely useful for smart grid technology are those that are unsupervised and focused on low frequency monitoring, which can be more reasonably expected to be available from multiple homes in a smart grid. Examples of this class of NILM method can be found in (Kim, Marwah, Arlitt, Lyon, & Han, 2011), (Kolter & Jaakkola, 2012), (Zhao, Stankovic, & Stankovic, 2016), among others. In this paper we use the REFIT dataset (Murray, Stankovic, & Stankovic, 2017), which includes submetering data for 20 houses in the Loughborough area of the UK in the period from July 2014 to July 2015, to evaluate the potential of NILM techniques in supporting smart grid analytics such as 1) When do the peaks in electricity consumption in the grid occur (aggregated across houses)? 2) Are there appliance loads during peak times with potential for deferred use (e.g. dishwashers, washers, dryers, etc.)? Can we recommend discount offers or time-of-use pricing programs to create incentives for customers to defer a portion of their electricity demand to times of lower overall demand? Which customers show greater potential and should be targeted first? 3) Can we generate more accurate demand forecasts at the grid level with the knowledge of individual consumer device make-up? We analyze the points above based on baseline truth from the REFIT dataset, as well as run the same analysis assuming inaccuracies on disaggregation results from NILM methodologies, in order to examine the consistency of the results. The goal is to evaluate how accurate NILM results need to be to enable this type of recommendation/customer relationship management, and what might be expected in terms of grid reliability or other return on investment metrics. Data description The REFIT dataset contains 20 houses, each with 11 metering points: 2 mains, which combined give the total apparent power drawn for the house, and 9 sub-metering points attached to 9 individual appliances within the house (appliances differ between houses). The instantaneous active power recordings are on average 8 seconds apart. We work with the cleaned dataset which re-aligns the sensors when appliances move within a house and imputes values (Murray, Stankovic, & Stankovic, 2017). Table 1 and Table 2 show additional information about the houses and appliances monitored. Houses 3, 11, and 21 from the REFIT dataset that affect the recording of total power consumption were removed from the analysis due to issues and do not appear on the tables. House 9 was also removed because it had too many gaps during our analysis period. 2018 Dell EMC Proven Professional Knowledge Sharing 4
House Occupancy 2 4 Dwelling Age 1975–1980 — Number of Appliances 35 15 1 2 4 5 6 7 8 10 12 13 15 2 4 2 4 2 4 3 4 1 1850–1899 1878 2005 1965–1974 1966 1919–1944 1991–1995 post 2002 1965–1974 33 44 49 25 35 31 26 28 19 16 17 18 19 6 3 2 4 1981–1990 mid 60s 1965–1974 1945–1964 48 22 34 26 20 2 1965–1974 39 Dwelling Type Size Detached Semidetached Detached Mid-terrace Detached Detached Detached Detached Detached Detached Semidetached Detached Detached Detached Semidetached Detached 4 bed 3 bed 4 bed 4 bed 4 bed 3 bed 2 bed 3 bed 3 bed 4 bed 3 bed 5 bed 3 bed 3 bed 3 bed 3 bed Table 1: Information about houses from the REFIT dataset. Source: (Murray, Stankovic, & Stankovic, 2017) 2018 Dell EMC Proven Professional Knowledge Sharing 5
Television Hi-Fi FridgeFreezer Fridge Freezer Microwave Cooker Hood Kettle Toaster Misc Kitchen Washing Machine Washer Dryer Tumble Dryer Dishwasher Computer Router Elec Heater Lamp Misc 1 X 2 X X X X 2X X X X X X X 4 X 5 X X X X X X X 2X 6 X House 8 10 X X X X X X X X X X X 7 X X X 2X X X X X X X 2X X X X X 2X X 12 13 15 16 17 18 19 2X 2X X X X X X X X X 2X X X X X X X X X X X X X 20 X Total X X X X X X X X X X X X X 7 13 13 1 12 7 3 16 X X X X X X X X X X X X X X X X X 2X X X X X X X X X X X X X X X X X X X X X X 2X X X X 18 2 10 1 8 11 12 0 3 1 2 Table 2 Appliances monitored in each house from the REFIT dataset. Source: (Murray, Stankovic, & Stankovic, 2017), Table 3. Since our objective is to simulate a small grid by assuming the REFIT houses are connected to the same subnetwork we focused our analyses on a 3-month time frame where data from all houses are available, from April to June 2014. To examine energy consumption over time for each home and for the entire grid, we initially use the power readings within each 15-minute block to compute energy consumption within that 15-minute block. For shortness we illustrate the energy consumption calculation in 1-minute blocks on Table 3. The same logic was used to compute energy consumption in 15-minute blocks. 2018 Dell EMC Proven Professional Knowledge Sharing 6
4/1/2014 0:02:00 4/1/2014 0:02:01 4/1/2014 0:02:14 4/1/2014 0:02:15 4/1/2014 0:02:19 4/1/2014 0:02:28 4/1/2014 0:02:30 4/1/2014 0:02:33 4/1/2014 0:02:44 4/1/2014 0:02:47 4/1/2014 0:02:58 4/1/2014 0:03:00 Aggregate power read (W) No Read 180 180 180 177 180 180 180 189 189 183 No Read 4/1/2014 0:03:01 177 Read timestamp Total Time delta (seconds) Energy consumption (Wh) 1 0.05 13 0.65 1 0.05 4 0.20 9 0.45 2 0.10 3 0.15 11 0.58 3 0.16 11 0.56 2 0.10* * computed assuming the power for the last 2 seconds of the 1-minute block is 177 W (read at 0:03:01) 60 seconds 3.04 Wh Table 3: Example of energy consumption calculation in a 1-minute block, given apparent power reads within that minute. We later leverage this initial 15-minute aggregation to produce hourly, daily, weekly, and monthly aggregations to examine consumption patterns. There are many gaps in the data that need to be filled to allow for an overall estimation of grid-level power consumption at any given time. Some gaps are short, lasting just a few minutes, others are long, lasting several days (up to 2 months in some cases). We handle these two types of gaps differently. Shorter gaps (lasting less than 15 minutes), are filled by the energy consumption calculation illustrated in Table 3. For longer gaps, we rely on averaging the closest prior and subsequent non-null data points for the same house, appliance, time of day, and day of week. Data exploration: aggregate energy consumption In our analysis we assume the 17 selected houses from the REFIT dataset are connected to the same subnetwork, forming a small grid. Although the REFIT houses were not chosen to fall on the same subnetwork, they were all located at the same region of England and subject to the same weather during the monitoring period, making the assumption of them being on a same subnetwork realistic in that respect. To examine patterns in grid-level energy consumption, we add the total energy consumed by all houses throughout the analysis period. 2018 Dell EMC Proven Professional Knowledge Sharing 7
Figure 1 shows the average grid-level energy consumption in kWh, aggregated by day of week and hour of day, for the 12-week period between April 6, 2014 – June 28, 2014, along with the standard error margins. To ensure our data imputation process doesn’t create artificial seasonality patterns, we removed periods with imputed reads from analysis. This means that 1-hour time periods for which at least one of the houses had a gap longer than 15 minutes in the power reads was not considered when computing average energy consumptions. We notice some patterns: Average energy consumption is higher on weekends than on weekdays (which is expected in cases where occupants must leave the house during weekdays for work and/or school obligations). Although patterns differ a bit from one weekday to the other, in general energy consumption peaks during a 5-hour block in the evenings from about 4PM to 9PM, with a smaller peak in the mornings, in the 3-hour block from about 5AM to 8AM (consistent with typical work/school schedules). Weekends also show two energy peaks, with an evening peak from about 4PM to 9PM (similar to weekdays), and a morning peak from about 7AM to 11AM (a bit later than weekday morning peaks). But energy consumption remains high in the middle of the day as well (albeit lower than during the mornings and evenings.) Figure 1: Average grid-level energy consumption in kWh, aggregated by day of week and hour of day, for the 12week period between April 6, 2014 – June 28, 2014. Based on these patterns, we define peak and off-peak hours as follows: Peak hours: 5 - 8AM and 4 – 9PM on weekdays, 7AM - 9PM on weekends. Off-peak hours: 12 - 5AM, 9AM – 4PM, and 9PM – 12AM on weekdays, 12 - 7AM and 9PM – 12AM on weekends. Although the peak-hours analysis can also be done by day of the week, we chose to focus only on the broader weekday/weekend distinction when evaluating potential for load deferral. This would result in energy tariff rules that are simpler and easier for customers to remember. 2018 Dell EMC Proven Professional Knowledge Sharing 8
Peak-period loads Periods of peak consumption can bring increased risk to the grid, so we aim to find opportunities for demand smoothing through incentives to change customer behavior. In this next section, we take a closer look into energy consumption by different houses and appliances during grid-level consumption peaks, singling out loads that customers could potentially defer to periods of lower grid-level demand. Out of the appliances listed on Table 2, washing machine, washer dryer, tumble dryer, and dishwasher stand out as potentially deferrable without having a big impact on lifestyle (no need to change meals or leisure schedules, for example), so we look for opportunities to defer these loads to off-peak hours. Figure 2 shows the total energy consumption (kWh) from April 6, 2014 – June 28, 2014 broken down by appliance during peak and off-peak periods, for Houses 5, 6, and 7. Given there are only 9 sub-metering points in each house, a good portion of the energy consumed cannot be attributed to an appliance and is marked as unassigned. We can notice several opportunities for load deferral across the houses. Houses 5 and 7 are good examples of load deferral candidates. They are among the top energy consumers in the grid, have higher total consumption during peak periods than during off-peak periods, and considerable (deferrable) consumption by dishwasher, tumble dryer, and washing machine during peak periods. Figure 2: Total energy consumption (kWh) from April 6, 2014 – June 28, 2014 by appliance during peak and off-peak periods, for Houses 5, 6, and 7. Table 4 shows a summary of total and deferrable (dishwasher/washer/dryer) loads by house, during peak and off-peak periods. In this table we also compute energy costs per house assuming two different tariff schemes: a flat tariff of 0.2 per kWh, or a variable tariff of 0.35 per kWh in peak periods and 0.05 per kWh in off-peak periods. When computing the energy cost for a house under flat tariff, we assume the original load distribution between peak and off-peak periods is maintained, since the customer would have no incentive to change habits. Under variable tariff, we examine two scenarios: 1) the customer changes tariff scheme, but does not change behavior (no load deferral) or, 2) the customer changes tariff scheme and defers the deferrable load (dishwasher/washer/dryer) to off-peak periods. In the latter scenario, it is assumed that the customers’ behavior changes are incentivized by the potential 2018 Dell EMC Proven Professional Knowledge Sharing 9
savings in the variable tariff scheme, but that they are also limited to deferral of loads that do not have a big impact on leisure/meals/sleeping habits (i.e. only the previously deferrable load is moved). In reality, customer behavior is difficult to predict. Some customers might place a higher value on maintaining their current schedules and decide to forfeit the potential savings from load deferral, even for tasks such as washing dishes and clothes. Others might be more cost-sensitive and decide to defer even more of their total load to off-peak periods, going beyond dish/clothing washing. These nuances in behavior should be examined through pilot programs in limited geographical regions and customer surveys, to more accurately forecast the effect tariff incentives may have in the overall population behavior. In the absence of supporting data on behavioral change, we aim to strike a balance in this paper by making the (strong) assumption that the value placed on keeping dish/clothing washing schedules is 0, so that customers would take advantage of any changes that bring savings (regardless of the amount), and the (conservative) assumption that dish/clothing washing are the only activities that customers would be willing to defer to off-peak periods. Examining Table 4, we notice that Houses 5, 7, and 8 would benefit the most from a change to variable tariff (higher cost savings), but Houses 5, 7, and 10 are the ones with higher deferrable loads during peak periods (would bring more benefit to the grid if behavior changed). Note that Houses 12 and 13 do not have enough deferrable load during peak periods to benefit from a variable tariff scheme, so in our simulation we assume they would not switch tariffs nor change behavior. Given our assumptions, making the variable tariff scheme available to this subnetwork would have the following effects from the utility’s perspective: Deferral of 847 kWh from peak to off-peak periods o Total peak period load changes from 7,275.3 kWh to 6,428.2 kWh o Total off-peak load changes from 6,796.0 kWh to 7,643 kWh Decrease in revenue from 2,814.2 to 2,579.3 due to discounts given for customers using more energy during off-peak periods. The demand smoothing gains from deferring 847 kWh from (peak) periods where energy consumption from other sources is still high, and hence reduces the chances of needing higher-cost energy generation and/or suffering outages, would need to be balanced against the revenue reduction of 235 due to tariff incentives. Given our assumptions that any cost savings would trigger costumer behavior changes, the utility could make the tariff change much less extreme, 0.25/0.15 instead of 0.35/0.05 for example, and still see the same amount of load deferral from peak periods while losing only 78.3 in revenue. In reality, higher cost savings are likely to trigger more significant customer behavior changes overall and promote more load deferrals. This tradeoff can be incorporated into the analysis once the relationship between potential cost savings and propensity to defer loads is better understood for different customer segments (through monitoring well designed surveys and pilot programs). 2018 Dell EMC Proven Professional Knowledge Sharing 10
Total load (kWh) Off% of total load Total energy Total energy % of total % of total Total Total energy Potential Likely to Peak during peak cost under cost load load during energy cost under savings switch Periods periods if all variable assuming during peak cost variable tariff from tariffs an deferrable load tariff with cheaper Total peak periods that under flat without load switching change is moved to offload deferral option is load periods is deferrable tariff ( ) deferral ( ) tariffs ( ) behavior? peak ( ) selected ( ) (kWh) 308.5 462.5 329.3 776.6 426.2 521.4 479.5 719.1 378.6 648.0 278.2 614.4 309.7 425.8 264.1 333.5 7275.3 376.2 304.7 381.2 562.2 503.7 412.4 752.5 624.0 279.1 395.8 288.5 558.3 298.0 463.6 270.1 325.8 6796.0 Peak Periods house 1 2 4 5 6 7 8 10 12 13 15 16 17 18 19 20 Total Deferrable Nonload Deferrable (kWh) load (kWh) 11.6 92.4 23.8 189.4 14.2 171.0 34.8 112.9 33.1 83.7 43.4 81.0 8.9 36.4 3.4 23.7 963.9 296.9 370.1 305.4 587.3 412.0 350.4 444.6 606.2 345.5 564.3 234.8 533.3 300.8 389.4 260.6 309.8 6311.4 45% 60% 46% 58% 46% 56% 39% 54% 58% 62% 49% 52% 51% 48% 49% 51% 1.7% 12.0% 3.4% 14.1% 1.5% 18.3% 2.8% 8.4% 5.0% 8.0% 7.7% 6.9% 1.5% 4.1% 0.6% 3.6% 43.4% 48.2% 43.0% 43.9% 44.3% 37.5% 36.1% 45.1% 52.5% 54.1% 41.4% 45.5% 49.5% 43.8% 48.8% 47.0% 136.9 153.4 142.1 267.8 186.0 186.8 246.4 268.6 131.5 208.8 113.3 234.5 121.5 177.9 106.8 131.9 2814.2 126.8 177.1 134.3 299.9 174.3 203.1 205.4 282.9 146.5 246.6 111.8 242.9 123.3 172.2 105.9 133.0 2886.1 123.3 149.4 127.2 243.1 170.1 151.8 195.0 249.0 136.5 221.5 98.8 218.6 120.6 161.3 104.9 125.9 2597.0 13.6 4.1 14.9 24.6 15.9 35.0 51.4 19.6 -5.0 -12.7 14.6 15.9 0.9 16.6 1.9 6.0 217.3 yes yes yes yes yes yes yes yes no no yes yes yes yes yes yes 123.3 149.4 127.2 243.1 170.1 151.8 195.0 249.0 131.5 208.8 98.8 218.6 120.6 161.3 104.9 125.9 2579.3 Table 4: Summary of total and potentially deferrable energy consumption, and costs under different tariff schemes, for all houses. Assumes there is sub-metering in place. The load deferral and cost savings analysis in Table 4 assumes that we have direct measures of the energy consumed by dishwashers, washers, and dryers. This is unlikely to be true for the majority of houses. In reality, these loads would be disaggregated through some sort of NILM methodology, which would produce estimates subject to accuracy limitations. Table 5 illustrates the potential impact of disaggregation errors in load deferral and revenue loss estimates for the utility, as well as in estimates of cost savings for customers. It assumes there is no sub-metering in place, and that the NILM methodology overestimates deferrable loads by 20% (total peak and off-peak loads remain the same as in Table 4, since they are measured by the main meter, but the breakdown between deferrable and nondeferrable peak load changes). Since under our assumptions the inflated deferrable load estimates would not cause a change in customer behavior patterns, the estimated load deferral would be inflated by 20% as well (to 1016.4 kWh). The inflation on estimated revenue loss would be around 21% (to 285.8). Table 6 shows the impact that varying deferrable load disaggregation errors would have in the accuracy of estimations for total deferred load and total revenue loss, under a variable-tariff based demand response program. Understanding how disaggregation inaccuracies impact estimated results of demand response programs is essential to the program design phase, ensuring that decisions are made taking into consideration the appropriate margin of error in estimated program impact to grid operation. 2018 Dell EMC Proven Professional Knowledge Sharing 11
Peak Periods NonDeferrable Total Deferrable load load load (kWh) (kWh) (kWh) 1 13.9 294.5 308.5 2 110.9 351.6 462.5 4 28.6 300.7 329.3 5 227.2 549.4 776.6 6 17.0 409.1 426.2 7 205.2 316.2 521.4 8 41.8 437.7 479.5 10 135.4 583.7 719.1 12 39.7 338.8 378.6 13 100.5 547.5 648.0 15 52.1 226.1 278.2 16 97.2 517.1 614.4 17 10.6 299.1 309.7 18 43.7 382.1 425.8 19 4.1 259.9 264.1 20 28.5 305.0 333.5 Total 1156.7 6118.6 7275.3 house Off% of total % of total load Total energy Peak % of total Total Total energy Total energy Potential Likely to load during during peak cost Periods load energy cost under cost under savings switch peak periods if all assuming during cost variable tariff variable tariff from tariffs an periods deferrable cheaper Total peak under flat without load with load switching change that is load is moved option is load periods tariff ( ) deferral ( ) deferral ( ) tariffs ( ) behavior? deferrable to off-peak selected ( ) (kWh) 376.2 304.7 381.2 562.2 503.7 412.4 752.5 624.0 279.1 395.8 288.5 558.3 298.0 463.6 270.1 325.8 6796.0 45% 60% 46% 58% 46% 56% 39% 54% 58% 62% 49% 52% 51% 48% 49% 51% 2.0% 14.5% 4.0% 17.0% 1.8% 22.0% 3.4% 10.1% 6.0% 9.6% 9.2% 8.3% 1.8% 4.9% 0.8% 4.3% 43.0% 45.8% 42.3% 41.0% 44.0% 33.9% 35.5% 43.5% 51.5% 52.5% 39.9% 44.1% 49.2% 43.0% 48.7% 46.3% 136.9 153.4 142.1 267.8 186.0 186.8 246.4 268.6 131.5 208.8 113.3 234.5 121.5 177.9 106.8 131.9 2814.2 126.8 177.1 134.3 299.9 174.3 203.1 205.4 282.9 146.5 246.6 111.8 242.9 123.3 172.2 105.9 133.0 2886.1 122.6 143.8 125.7 231.8 169.2 141.5 192.9 242.3 134.5 216.4 96.2 213.8 120.1 159.1 104.7 124.5 2539.1 14.3 9.6 16.4 36.0 16.7 45.2 53.5 26.4 -3.0 -7.7 17.2 20.8 1.4 18.8 2.1 7.4 275.1 yes yes yes yes yes yes yes yes no no yes yes yes yes yes yes 122.6 143.8 125.7 231.8 169.2 141.5 192.9 242.3 131.5 208.8 96.2 213.8 120.1 159.1 104.7 124.5 2528.5 Table 5: Summary of total and potentially deferrable energy consumption, and costs under different tariff schemes for all houses. Assumes there is no sub-metering and the NILM methodology overestimates deferrable loads by 20%. Deferrable Estimated load Deferred disaggregation Load error 278.21 -50% 447.47 -40% 528.25 -30% 603.72 -20% 762.34 -10% 847.04 0% 931.74 10% 1016.45 20% 1101.15 30% 1185.85 40% 1270.56 50% Estimated Revenue Loss Estimated/Actual Estimated/Actual Deferred Load Revenue Loss 121.91 140.52 163.00 185.64 209.55 234.96 260.38 285.79 311.20 336.61 362.02 33% 53% 62% 71% 90% 100% 110% 120% 130% 140% 150% 52% 60% 69% 79% 89% 100% 111% 122% 132% 143% 154% Table 6: Impact of deferrable load disaggregation errors in accuracy of total deferred load and revenue loss estimations under a variable-tariff based demand response program. Demand forecasting We now turn to electric grid demand forecasting, or predicting future energy use on some relevant time interval in advance. Better control over those forecasts translate into better control over incentives for end customers. We argue that results from an energy disaggregation implementation can increase the effectiveness of demand reduction incentives. First, in providing more accurate demand forecasts by 2018 Dell EMC Proven Professional Knowledge Sharing 12
including the low-level disaggregated appliances on the end customers’ level. Second, in providing a more targeted offer which can prove more reliable in predicting the response to the offer. Using our simulated grid aggregated to the hour, we build a demand forecasting model in two cases. First, in the case where only aggregate data is known and used for the model. Second, when disaggregation information is known. For the comparison, we use a simple demand forecasting model based on ARMA to demonstrate the two cases. Using the 3-month period we have chosen, weekly training data sets and single-day forecasting is chosen. This provides at least 11 different weeks of model application for comparison. ARMA Model The ARMA model incorporates autocorrelations in a regression approach, parametrized by maximum correlation factors 𝑝 and 𝑞. It is well known that electricity usage for residential customers has daily, weekly, and yearly patterns, and any demand forecasting model must incorporate these patterns as a first step. As is typical for the model, we compute the autocorrelations. For the 3-month period, Figure 3 shows the autocorrelation factor (ACF) for each house in the dataset. Vertical lighter bands in the image correspond to relatively strong correlations for each multiple of 24 hour lags across houses. For some houses, the correlation increases slightly at the 7-day lag, 168 hours. If the data is differenced by 24 hours, the correlation decreases significantly. Figure 3: Heatmap of each house’s ACF for the three-month period. For comparison, Figure 4 shows the partial autocorrelation factor (PACF) for the same data. Most of the correlations here are not significant, except for one peak for house 8 at 24 hour lag. Most houses only have a significant correlation for a one-hour lag. 2018 Dell EMC Proven Professional Knowledge Sharing 13
Figure 4: Heatmap of the PACF for the three month period. We set up our model to predict the future day, given a week of data. Due to the 24-hour correlation, we difference the consumpt
Smart meters can be deployed as a single meter for a customer or in complex situations in order to measure specific devices. However, the field of Non-Intrusive Load Monitoring (NILM) aims to disaggregate high-level aggregate measurements to contributions of individual appliances, based on a single metering point (smart meter at main breaker
monitoring energy consumption: the Intrusive Load Monitoring (ILM) approach, which use one sensor for each appliance, and the Non-Intrusive Load Monitoring (NILM) approach which aims at estimating the load consumption from a unique overall current and voltage measurement. Figure 1 presents the difference between the two approaches.
Bruksanvisning för bilstereo . Bruksanvisning for bilstereo . Instrukcja obsługi samochodowego odtwarzacza stereo . Operating Instructions for Car Stereo . 610-104 . SV . Bruksanvisning i original
10 tips och tricks för att lyckas med ert sap-projekt 20 SAPSANYTT 2/2015 De flesta projektledare känner säkert till Cobb’s paradox. Martin Cobb verkade som CIO för sekretariatet för Treasury Board of Canada 1995 då han ställde frågan
service i Norge och Finland drivs inom ramen för ett enskilt företag (NRK. 1 och Yleisradio), fin ns det i Sverige tre: Ett för tv (Sveriges Television , SVT ), ett för radio (Sveriges Radio , SR ) och ett för utbildnings program (Sveriges Utbildningsradio, UR, vilket till följd av sin begränsade storlek inte återfinns bland de 25 största
Hotell För hotell anges de tre klasserna A/B, C och D. Det betyder att den "normala" standarden C är acceptabel men att motiven för en högre standard är starka. Ljudklass C motsvarar de tidigare normkraven för hotell, ljudklass A/B motsvarar kraven för moderna hotell med hög standard och ljudklass D kan användas vid
LÄS NOGGRANT FÖLJANDE VILLKOR FÖR APPLE DEVELOPER PROGRAM LICENCE . Apple Developer Program License Agreement Syfte Du vill använda Apple-mjukvara (enligt definitionen nedan) för att utveckla en eller flera Applikationer (enligt definitionen nedan) för Apple-märkta produkter. . Applikationer som utvecklas för iOS-produkter, Apple .
Non-Intrusive Pig Signaller Magnetic Topside Lithium Operating Manual The 4001D MAGSIG is a fully ATEX and IECEx certified non-intrusive pig signaller which quickly and accurately detects, signals and logs the passage of magnetic pigs at critical points along a pipeline both on land and offshore. Online Electronics Limited 44 (0) 1224 714 714
modern kWh-meter Hannu Pihala VTT Energy This licentiate thesis has been submitted for official examination for the degree of Licentiate in Technology in Espoo, May 1998. . Pihala, Hannu. Non-intrusive appliance load monitoring system based on a modern kWh-meter. Espoo 1998, Technical Research Centre of Finland, VTT Publications 356. 68 p .
