Context Monitoring Optimization for Demand Response

Context Monitoring Optimization for Demand Response

DR methods usually reflect national policies and are triggered by different events; however they are all based on one of two  paradigms: Indirect Load Control, where the consumer reacts to external stimuli and adjusts his consumption (e.g. new price) or Direct Load Control, where specific devices are automatically adjusted according to specific agreement between the consumer and the provider.

Although effective in theory, in practice both approaches suffer from significant drawbacks: direct control disregards user behavior and habits while indirect approaches necessitate considerable interaction with the customer, thus creating discomfort.

It is evident that a new approach is required, one that combines the salient features of both direct and indirect DR paradigms so as to appropriately adjust energy consumption while minimizing end user discomfort and intrusion. This implies the design and implementation of automated tools that use explicit and/or implicit information and limit direct interaction with the users. In turn, this necessitates the dynamic definition of a user profile based on the aggregation of various information sources (e.g. questionnaires, building sensors, and actuators), the analysis of user-driven events as well as the reaction of the user to external -system generated- stimuli.

This envisaged paradigm shift calls for a personalized framework that operates within the consumer’s residence. The profiling mechanism should be deployed in a low end controller device (e.g. Raspberry Pi – [3], Soekris – [2] or similar embedded PC), affordable and handy and not larger than an ADSL router, that constantly monitors energy consumption. 
Data acquisition performed by the monitoring process is an essential part of this profiling mechanism. The rate of sampling is a crucial factor since it is related to: 
i)    the successful/unsuccessful detection of events, 
ii)    the processing power needed to perform the sampling and, 
iii)    the energy that the device and the sensor nodes consume during such actions.

In order to address these issues we designed a simple and efficient mechanism that dynamically adapts the sampling rate of the monitoring procedure.  The algorithm attempts to kill two birds with one stone; on one hand, in the absence of events minimize the number of unnecessary monitoring actions, while on the other hand detect –ideally all– events when they appear.  

From a high level, methodological point of view, upon initiation, the algorithm searches for events scanning at the highest possible frequency (e.g. one scan per designated time). In the absence of events, its scanning rate is progressively reduced until a predefined, low threshold is reached. When an event is identified the algorithm is re-calibrated to the highest possible scanning frequency and keeps measuring at this frequency until no event is monitored. The same process is repeated until all events have been identified.  

The merits of the mechanism have been quantified by means of an analytical model. We model the procedure as a Markov chain, each state of which essentially adjusts the scanning frequency of the monitoring procedure. We assumed that a scanning/sensing/monitoring attempt is successful if it identifies an event, and unsuccessful otherwise. Each state of the Markov chain is characterized by a monitoring frequency. If the algorithm detects the occurrence of an event, it transitions to state Sk in which it operates at the highest possible monitoring frequency. If it does not detect an event, it transitions to a ‘lower’ state (towards state S1), in which sensing is sparser. Transition probabilities for any state Si are Pi and 1- Pi where the former denotes the probability of transition from Si to Sk and the latter from Si to Si-1.

Further analysis indicates that the key factor that should be taken into account is the density of events d, i.e. the frequency of events occurrence during a day. After identifying d, then we can precisely calculate the length of the chain and thus calibrate the algorithm so as to detect the highest amount of events with the minimum possible number of loops.
The added value of the aforedescribed algorithmic solution is twofold. First and foremost, it achieves a significant reduction in the CPU load of the monitoring device. CPU-load reduction is extremely important in our case since the embedded devices that accommodate the profiling framework are equipped with a single CPU (i.e. a Raspberry Pi comes with a 700MHz ARM processor) thus performance optimization is a strict pre-requisite. It is worth mentioning that theoretic analysis indicates that in eventful conditions, we can monitor more than 70% of events by performing 50% less loops than the always-on-case. Additionally, the generic nature of the mechanism enables its application on any case that can be reduced to an event-detection use case. In simpler terms, such a solution will help establish an affordable, easy-to-use, non – intrusive method of Demand Response among household consumers, making DS easier all around, more efficient and even more cost-effective in dealing with peak demands.

Our work now focuses on developing a learning scheme so as to enable the algorithm to take into consideration past information and seasonality in order to autonomously calibrate its input parameters without requiring any human intervention. Obviously, each human is unique, thus a one-size-fit-all solution is not adequate. Our plan is to design an algorithm that will be adaptable to the end user, learn from him and eventually, after a period of time, identify and adapt to its behavioral (in terms of events generation) pattern.

[1]    Balijepalli, Murthy; Pradhan, Khaparde (2011). “Review of Demand Response under Smart Grid Paradigm”. IEEE PES Innovative Smart Grid Technologies.
[2]    Soekris Engineering:
[3]    Raspberry Pi:

Panagis Magdalinos
Data scientist