AC Cooling Time Prediction Using Common Representation Model

Ramasamy Kannan,Manasij Sur Roy

doi:10.1109/access.2020.3009467

Ramasamy Kannan, Manasij Sur Roy

Open Access

https://doi.org/10.1109/access.2020.3009467

Copy DOI

Abstract

The data generated by Air Conditioner (AC) consists mainly of sensor and control data. This paper will use the data generated from 53,528 ACs to predict the AC cooling time. The cooling time is the time taken by the AC to cool to a desired temperature. We have observed certain important issues in the data gathered from ACs deployed in dynamic real world environments. Poor prediction accuracies are observed for about 76% of the total ACs due to the lack of data regarding the device behavior, AC settings selection behavior and environmental conditions. During the AC operation, it is observed that the user selects only a small subset of the various combinations of the overall possible settings. Due to unavailability of data, Machine Learning (ML) models cannot be generated for new ACs. This leads to a cold start problem. This paper proposes a common AC prediction model that is generated through data shared from multiple connected ACs. Additionally an Auxiliary Task Learning (ATL) based deep learning model will be used for improving prediction accuracy. The proposed solution provides prediction capabilities for all ACs, compared to 24% of ACs supported by individual prediction models.

Highlights

An Air Conditioner (AC) has many sensors and control settings that generate data during the operation of the device
We have shown that even for the best Machine Learning (ML) model generated from the standalone AC data, more than 76% of the ACs have R2 less than zero
This indicates that the data generated by standalone ACs are not sufficient for designing prediction models

Summary

Introduction

An Air Conditioner (AC) has many sensors and control settings that generate data during the operation of the device. Each AC has multiple controls/settings that helps user customize its operations. In this paper we improve AC representation learning through data collected from multiple ACs and use it to predict AC cooling time. The generated common model will work on all ACs including the majority of the cases where there is insufficient data to generate prediction model for individual ACs. A survey of algorithm based intelligent HVAC (Heating, Ventilation and Air Conditioning) control systems for predicting outdoor and indoor temperature, user presence, thermal preference and energy reduction are discussed in [1]. AI solutions to predict building cooling loads in order to reduce AC energy are discussed in [6], [7]. An algorithm to control the air flow volume in Air Conditioners is described in [8]

Objectives

Methods

Results

Conclusion