The proposed energy consumption saver is an innovative technology that aims to increase the efficiency of energy consumption in residential buildings, production and commercial facilities, and other types of structures. It integrates an artificial intelligence (AI) system that reads energy consumption data autonomously through power points (electrical outlets with wireless adapters attached to them), analyses the collected information, calculates optimal consumption levels, and informs users about a need to adapt consumption behaviours.
This report will review the major elements contributing to the effective functioning of the AI-based technology, including datasets, pre-processing factors, and machine learning mechanisms. Additionally, potential sources of error that may arise from data sources will be discussed, along with the primary ethical and security implications of the AI system.
Machine Learning Model: Reinforcement Learning
For data analysis, the energy consumption saver will utilise the Reinforcement Learning (RL) algorithm. The RL model is concerned with “how a learning agent learns what to do (action) in a given situation (state) by interacting with an environment to maximize or minimize numerical returns (rewards)” (Kim & Lim, 2018, p. 2). A schematic depiction of basic RL mechanisms is given in Figure 1.
In formal terms, the primary problem of this learning model is to optimise a reward function R within a certain location or state in an environment x. The objective of RL is to define the variable of location/state x* that affords the maximum reward R(x*) (Sutton & Barto, 2018). Compared to Deep Learning that implies learning from a training set and then implementing the same mechanism on newly gathered data, RL is usually not provided with any training examples.
However, for the proposed energy consumption monitoring technology, RL will be given an initial system condition based on the consumption and expense preferences of power consumers in a building. Two conditioning variants, cost-preferred and comfort-preferred, will be available, and a user will be able to find the most suitable balance between the two solutions.
The main distinguishing feature of RL – a high degree of dynamism – will be preserved in the proposed AI technology. The term “dynamism” implies here that the RL system adjusts its actions depending on continual feedback in order to maximise rewards, that is, the best-expected outcomes associated with power use (Marr, 2018). In other words, the AI technology will try different actions aimed to increase the efficiency of energy consumption and will evaluate whether a certain action helps to achieve a more favourable change. Afterwards, it will reinforce the actions that work best and, in this way, the computer will engage in an ongoing, autonomous modification of its algorithms through the process of trial and error until a decision delivering the most advantageous results is reached.
Another important feature of RL is a long-term memory that allows for bypassing exploratory search and minimising the need to repeat actions that were unsuccessful before. Over some time, the system builds a substantial library of states (Dooraki & Lee, 2018). In the proposed AI project, these states will include optimal energy consumption outcomes, including the overall amount of energy utilised and electricity costs incurred, as well as related user behaviours. It is worth noting that search is a continual process in the RL model, yet its results are arranged in a way that reduces search efforts or entirely eliminates them over time and as significant experience is gained.
Potential Datasets and Pre-processing Considerations
Two major datasets will be required for the determination of rewards: energy demand and price. The first category entails the actual amount of energy consumed in a household or an organisation per hour (kWh). The second category denotes the cost of energy transactions ($/kWh) and will be determined based on the current information available in the energy market. The data will be collected through multiple power points located across a building or a particular residential area, and the two information categories associated with these specific locations will be calculated separately. The system will take into account unusual user behaviours (namely, on/off adapter statuses, timeframes of energy consumption, and so forth) and will notify users when adjustments in consumption are required.
In order to calculate information relevant to cost-preferred and comfort-preferred power consumption, it is essential to develop a unified algorithm framework that considers both of these variables at the pre-processing stage. As stated by Ouyang (2018), “having multiple frames as input would take up much more processing power and time to train” (para. 4). Thus the input approach that uses two sequential frames may be regarded as ineffective. It will be necessary to elucidate the outcome differences between the solutions, maximising either comfort or cost rewards. Consequently, a differential frame/graph combining both of these necessary components will be created. This will help the RL model to make informed decisions and to find the right balance between cost efficiency and comfort.
Sources of Error and Bias
Every automated decision-making process is inherently associated with a risk of bias. For instance, energy consumption activities involve an aspect of moral judgement linked to a desire to minimise environmental impact. In addition, the preference of energy consumption patterns includes an element of user choice, and it is unlikely that one would want to contribute to the improvement of environmental sustainability by sacrificing a great deal of comfort.
Thus automation bias may arise if one variable is favoured over another. This type of bias is defined as capture bias and is linked to the way data are acquired “in terms of the used device and of the collector preferences” (Tommasi, Patricia, Caputi, & Tuytelaars, 2017, p. 38). The provision of appropriate initial system conditions can help to resolve this problem (Bajwa, 2018). The optimisation of parameters through preliminary training will allow for maximising the model’s accuracy.
As a dynamic model, RL is also prone to overfitting errors. Flexible procedures can sometimes “be picking up some patterns that are just caused by random chance rather than by true properties” of the function (Gupta, 2017, para. 2). It is possible to minimise this problem by mainly focusing on data inference processes, rather than the prediction of results, and by using a simpler algorithm.
Ethical, Social, and Security Implications
The proposed technology has both positive and negative implications for its users and society in general. First, such smart devices as the automated energy consumption saver provide flexibility in the monitoring of power consumption, and they support relevant lifestyle changes (Bhati, Hansen, & Chan, 2017; Forbes Technology Council, 2018).
Without them and without the numerical and visual evidence they offer to users, it may be more difficult to alter lifestyles because human decision making is linked to multiple cognitive and psychological biases, and people are frequently unable to simultaneously consider the multitude of diverse factors needed to make the right choices (Manninno et al., 2015). While individual and business users may be primarily motivated to utilise AI for energy saving because of the greater cost-efficacy (Berger, 2018), its potential impact on overall energy conservation at the community, national, and global levels should not be undervalued as well.
As for possible negative implications, the main ones are security and privacy risks. Although the technology may not aim to capture personal and sensitive data as such, it still records and stores a lot of information about the way individuals conduct their everyday lives. It is clear that these data can be misused and abused. The main problem is that in the case of large, modern datasets, the collected information is frequently impossible to anonymise entirely (Montjoye, Farzanehfar, Hendrickx, & Rocher, 2017). In this way, technology users may be exposed to a risk of cybercrime.
Another potential but less immediate social ramification of the suggested technology is its effect on employment rates. AI-performed labour is usually cheaper than human labour, and technologies are typically more efficient in fulfilling mechanical and analytical tasks (Huang & Rust, 2018). Thus, the more AI technology is applied, the less need there may be for employers to hire employees. However, from a different perspective, the integration of automated energy consumption savers in buildings may foster the creation of new jobs because competent and technologically savvy workers will be required for the technical maintenance and control of the device.
References
Bajwa, A. (2018). What we talk about when we talk about bias (a guide for everyone). Medium. Web.
Berger, R. (2018). Artificial intelligence: A smart move for utilities. Web.
Bhati, A., Hansen, M., & Chan, C. M. (2017). Energy conservation through smart homes in a smart city: A lesson for Singapore households. Energy Policy, 104, 230-239.
Dooraki, A. R., & Lee, D. J. (2018). Memory-based reinforcement learning algorithm for autonomous exploration in unknown environment. International Journal of Advanced Robotic Systems, 15(3), 1-10.
Forbes Technology Council. (2018). 14 ways AI will benefit or harm society. Forbes. Web.
Gupta, P. (2017). Balancing bias and variance to control errors in machine learning. Towards Data Science. Web.
Huang, M.-H., & Rust, R. T. (2018). Artificial intelligence in service. Journal of Service Research, 21(2), 155-172.
Juvin, G. (2017). My journey into deep Q-learning with Keras and Gym. Medium. Web.
Kim, S., & Lim, H. (2018). Reinforcement learning based energy management algorithm for smart energy buildings. Energies, 11(8), 1-19.
Mannino, A., Althaus, D., Erhardt, J., Gloor, L., Hutter, A., & Metzinger, T. (2015). Artificial intelligence: Opportunities and risks. Web.
Marr, B. (2018). Artificial intelligence: What’s the difference between deep learning and reinforcement learning?Forbes. Web.
Montjoye, E. A., Farzanehfar, A., Hendrickx, J., & Rocher, L. (2017). Solving artificial intelligence’s privacy problem. Field Actions Science Reports, Special Issue(17), 80-83.
Ouyang, J. (2018). Reinforcement learning lane change algorithm (part 1). Good Audience. Web.
Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). Cambridge, MA: The MIT Press.
Tommasi, T., Patricia, N., Caputi, B., & Tuytelaars, T. (2017). A deeper look at dataset bias. In G. Csurka (Ed.), Domain adaptation in computer vision applications (pp. 37-58). Cham, Switzerland: Springer.