Viedās mājas apkures uzlabošana ar edge AI.

Abstract

Veikalos nopērkami viedie termostati ir viegls un salīdzinoši lēts risinājums kā uzlabot istabas temperatūras komfortu un ietaupīt enerģiju. Nest, Netatmo, Ecobee un Tado ir populārākie risinājumi tirgū, bet tie ir atkarīgi no mākoņa savienojuma, un ļauj kompānijām izmantot lietotāju datus. Lai atteiktos to risinājumiem, kam nepieciešams savienojums ar mākoni, tiek piedāvāts alternatīvs viedās mājas termostata risinājums, kas uzlabo esošos programmējamos termostatus, izmantojot mašīnmācīšanās algoritmus, lai aprēķinātu turpmākos apkures laikus un kontrolētu radiatorus. Tas izmanto lineārās regresijas algoritmu un ārējo temperatūru kā papildus mainīgo, lai laicīgi ieslēgtu radiatorus un norādītajā laikā sasniegtu mērķa temperatūru. Mašīnmācīšanās modelis tiek pārbaudīts teorētiski, izmantojot daļu no savāktajiem datiem priekš testēšanas, un praktiski, lietojot termostata programmu vienu nedēļu divās atsevišķās telpās. Testēšanas laikā tiek secināts, ka lineārās regresijas algoritms strādā gadījumos, kad rezultātus neietekmē nezināmi ārēji faktori; praktiskā pārbaude apstiprina šos rezultātus. Rezultāti liecina, ka ir lineārās regresijas modelis var tikt izmantots viedajiem termostatiem; tomēr, lai tas precīzāk strādātu, testēšanas datu kopā ir jāiekļauj visi ārējie faktori. Lai uzlabotu praktiskajā daļā izveidoto termostatu, ir ieteicams apkopot vairāk ievades datus sākotnējai modeļa apmācībai, kā arī veltīt vairāk laika praktiskai modeļa testēšanai. Bakalaura darbs ir uzrakstīts angļu valodā, tas sastāv no 52 lapaspusēm, 17 figūrām un 47 literatūras avotiem. Atslēgas vārdi: mākslīgais intelekts, viedā māja, mašīnmācīšanās, perifērdatošana, viedais termostats.

Bachelor’s thesis “Improving Smart Home Heating with Edge AI” explores smart thermostats as they have become increasingly popular during the past decade due to their low costs, easy installation, and features they present. They offer energy savings by optimizing heating schedules and stopping heating entirely when no one is home, as well as provide effortless temperature adjustments and schedule creations. The most popular brands on the market, such as Nest, Netatmo, Ecobee, and Tado, all have multiple solutions for different use cases and devices; nevertheless, they might not be the ideal choice. Their dependence on a cloud connection can cause potential security risks, making them vulnerable to attacks. Furthermore, relying on a third-party server does not guarantee that thermostats will always be up and running, and it is often unknown how the servers process and use the data they receive. To combat this, an alternative smart home thermostat solution is offered that improves on existing programmable thermostats by using a supervised machine learning algorithm on edge to calculate future heating times and control the thermostat. It uses a linear regression algorithm and outside temperature as an additional variable to predict heating time until the following schedule changes and turns on the radiators early to reach the target temperature at the specified time. This intelligent functionality improves homeowners' comfort as they can set the desired temperature when they, e.g., arrive home from work, and the thermostat decides when the heating should start to reach the target temperature at a specified time. The machine learning model is tested theoretically using part of the collected data as a testing set for accuracy evaluation in three rooms. It shows that the linear regression algorithm can correctly predict heating times in cases where no unknown external factors impact the results, such as independent, uncontrollable heating sources or varying room occupation. The thermostat works very well in one of the rooms, acceptably in another room, but due to external factors, it fails to predict heating times in the third room, which is why the practical testing is conducted only in two rooms. Practical testing involves using the program for one week in two separate rooms and confirms the theoretical results by showing the model’s ability to predict heating in some cases but not others. One of the rooms is heated by the smart thermostat as expected, and the room often reaches the target temperature; however, it either happens later than it should, and the room is overheated, or it does not start heating early enough to reach the target temperature. Due to warm weather during the testing week, heating is not required in the second room; thus, its results are not considered. It is also concluded that the model does not deal well with room overheating issues despite prior data refactoring. The results show that the linear regression model for machine learning smart thermostats is applicable; nevertheless, external factors must be included in the testing data set; otherwise, the results are unacceptable. The biggest constraint of the testing is the timeframe. Due to long heating cycles, additional real-world testing is required to finetune the model and make it better suited as an improvement to programmable thermostats. This could be achieved by software simulation, though it still takes time. Moreover, other machine learning algorithms, such as Naive Bayes or neural networks, could be explored for improved heating predictions as they have shown positive results in other heating solutions. To further enhance the on-the-edge smart thermostat, collecting more input data for initial model training is suggested, as well as considering overheating as a separate variable in the training dataset and spending more time on practical model testing. This would make the on-the-edge thermostat more practical and a valuable alternative to exi