Optimal energy allocation for households in generation-constrained off-grid microgrids

Abstract

For microgrids with limited generation capacity, allocating a daily
equal energy budget to each household is one way of ensuring that
all households are provided with sufficient energy to be used in a
day without compromising the daily operation of the power system.
With the same daily energy quota, households are given freedom on
how to spend energy according to their preferences and priorities.
Issues in this type of energy management scheme include 1) power
outage in households that use up all their energy allowance before
the scheduled replenishment and 2) unused energy allocation turned
to waste from households that are unable to consume the energy
allowance. Energy waste in terms of unused energy allocation of some
households can be beneficial to other households. The unused energy
can be distributed to other households that experience a power outage
and need more energy than the allocated. One approach to solving the
above issues is to frame the problem as an optimisation problem that
aims to minimise the energy wastage and maximise energy availability.
This research proposes an optimal energy allocation for households
connected to generation-constrained microgrids. The proposed optimal
energy allocation scheme has two main parts. First, the ideal
energy utilisation of each household is predicted using a multi-layer
perceptron (MLP); secondly, the optimal energy allocations for the
households based on their predicted utilisations are derived using
Karush-Kuhn-Tucker (KKT) optimality conditions. From the application
of the KKT conditions, a methodology for optimal energy allocation
that is adaptive to each household is proposed. The approach is
optimal as it minimises the energy wastage/deficit while maximising
energy availability to households, and adaptive because it uses the
household’s historical data and demographic information.
To support the development of the MLP-based forecast model, this
thesis implemented an energy monitoring system called Philippines
Micro-Off-Grids (PMOG) system to gather the actual historical energy
usage data in representative households from select villages in Cebu,
Philippines which have microgrids with limited generation capacity.
In the Philippines, there are 40 million people without access to electricity
[Och13] and microgrids are used to provide electricity access
to villages that are not accessible by the traditional grid. There are
three villages selected with two of them being off-grid communities
and one being grid-tied community. The energy data from PMOG
system serves as the baseline data for the development of the forecast
model and the optimal allocation scheme. A survey is also conducted
to gather the household demographic information that affects their
daily energy consumption.
This thesis presents an experimental method to determine the best
combination of hidden layers and neurons of the neural network along
with the input delay window in shaping the input variables that allows
the forecast model to generate the lowest possible root mean squared
error (RMSE). Since the optimal energy allocation is dependent on the
accuracy of MLP-based load forecast model, the right combination of
those design parameters of the neural network together with the delay
window used in shaping the inputs is crucial. These parameters are
considered as the main factors affecting the performance of the neural
network in forecasting.
Experimental results show that as the households’ demographic
information is included as input variables secondary to the historical
energy data and weather information, the performance of the neural
network improves significantly. The RMSE decreases from 92 W to 81
W, which represents a 12 % decrease for a neural network with three
hidden layers, 20 neurons and seven delays.
Given the limited generation capacity of the microgrid, the objective
is to minimise the squared difference between the ideal utilisation of
the household (which is estimated by the MLP-based forecast model)
and the (calculated) allocated energy. Results from the data from
the select villages show an aggregated unused or deficit energy per
household (for a day) from the existing equal allocation can be reduced
from 0.24 kWh to 0.11 kWh using the proposed dynamic/adaptive
allocation, which is about 54 % reduction in unused or deficit energy.
For 288 days, a total of 44 % reduction of energy wastage is achieved
with the proposed methodology when compared with equal allocation,
that is 112 kWh using equal energy allocation, and 62 kWh using the
proposed optimal energy allocation.
In summary, the proposed approach of allocating the daily energy
allowance of the household which is a hybrid approach using an MLPbased
forecast model and KKT optimality conditions minimises the
unused energy and enables households to maximise their energy usage
without compromising the minimum energy requirement of each
household in villages powered by microgrids with limited generation
capacity. By incorporating household profiles as inputs to the MLPbased
forecast model, prediction accuracy was improved by 12% in
terms of RMSE. From my experiments, employing MLP-based forecast
model ensures better forecasting performance than other techniques
such as Autoregressive Integrated Moving Average (ARIMA), Radial
Basis Function Network (RBFN) and Gaussian Process Regression
(GPR). The overall average accuracy for the MLP-based forecast model
is 91% with the highest accuracy of 93% for House 5 predictions, and
the lowest is 91% for House 3.
This approach is expected to work on households with similar
profiles connected to any off-grid power systems. Optimising the daily
energy quota will enable the village to maximise the usage of the
available energy with minimum wastage in terms of unused energy
quota. This approach will also lead the village to have a better payment
scheme based on their actual usage of electricity.

Date of Award	13 Oct 2022
Original language	English
Awarding Institution	Coventry University University of San Carlos
Sponsors	British Council & Science Education Institute - DOST
Supervisor	James Brusey (Supervisor), Elena Gaura (Supervisor) & Kojo Sarfo Gyamfi (Supervisor)