=Paper=
{{Paper
|id=Vol-2267/313-317-paper-59
|storemode=property
|title=Improving the efficiency of energy consumption in smart grids with application of artificial intellect
|pdfUrl=https://ceur-ws.org/Vol-2267/313-317-paper-59.pdf
|volume=Vol-2267
|authors=Eugene Yu. Shchetinin
}}
==Improving the efficiency of energy consumption in smart grids with application of artificial intellect==
https://ceur-ws.org/Vol-2267/313-317-paper-59.pdf
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
IMPROVING THE EFFICIENCY OF ENERGY
CONSUMPTION IN SMART GRIDS WITH APPLICATION OF
ARTIFICIAL INTELLECT
E. Yu. Shchetinin
Financial University under the Government of Russian Federation, Moscow, Russia
e-mail: riviera-molto@mail.ru
Energy saving accuracy estimates are important for development of energy efficient projects and for
demonstrating their cost-effectiveness. Increasingly, commercial buildings have an advanced
measurement infrastructure, which has led to the availability of high-frequency interval data. These
data can be used in a number of energy efficiency tasks, including anomaly detection, control and
optimization of heating, ventilation, and air cooling systems. All it makes possible the application of
artificial intellect methods and therefore leads to more accurate estimates of energy savings. In this
paper, we proposed the method for modeling the consumers energy profile based on the clustering
analysis, exactly, modified K-means algorithm. Extended computer simulations have been shown that
it improves the accuracy of energy consumption forecasts considerably.
Keywords: energy consumption, smart meter data, smart grids, forecasting, cluster analysis, gradient
boosting.
Β© 2018 Eugene Yu. Shchetinin
313
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
1. Introduction
The development of the intelligent networks in industry, finance and services creates new
opportunities for the development and application of effective methods of machine learning and data
analysis. Smart technologies for collecting, recording and monitoring data on energy consumption
create a huge amount of data of different nature. These data can be used for optimal network
management, improving the accuracy of the forecasting load, detection of abnormal effects of power
supply (peak load conditions), the formation of flexible price tariffs for different groups of consumers
[1, 2, 3]. One of the most important issues in this area is to predict the power load consumption as
accurately as possible. Consumption have a rather complicated stochastic structure, which are difficult
for modeling and prediction. Nevertheless, when different methods of aggregation are applied to the
group of consumers having similar statistical characteristics of time series of power consumption, it is
possible to count on considerable progress in the solution of objectives. However, in order to improve
the accuracy, our goal is to investigate the possible benefit of combining clustering procedures with
forecasting methods. We have used several known approaches such as Holt-Winters exponent
smoothing, ARIMA model, Support Vector Regression and some others. Also most effective machine
learning algorithms were applied, exactly, random forest, gradient boosting and bagging.
2. Machine learning algorithms applications for energy consumption
modeling
The problems of application of clustering methods to the time series of electricity
consumption are mainly in high dimension and high noise level of the data, which can be solved with
the use of machine learning methods [4,14]. Our method consists of three steps: the first one is to
normalize the data and calculate the energy consumption model for each consumer. In the future, the
study uses four different models based on the representation of time series, which serve as inputs to
the clustering method. The second stage consists of calculating the optimal number of clusters for the
given time series representation and the selected data training period. The third stage is clustering and
aggregation of consumption within clusters. For each cluster, the forecast model is trained and the
forecast for the next period is run. Then the forecasts are aggregated and compared with the real
consumption data. Next, we construct a forecast for received representations of the clusters using the
methods, that we will describied as follows.
2.1. Modeling of the energy consumption time series
The time series X is an ordered sequence of n real variables
πΏ = (ππ, ππ, β¦ , ππ), ππ β πΉ. (1)
The main reason for using time series presentation is a significant decrease in the dimension of the
analyzed data, respectively, reducing the required memory and computational complexity. Four
different model-based representation methods were chosen: (a) Robust Linear Model (RLM), (b)
Generalized Additive Model (GAM), (c) Holt-Winters Exponential Smoothing (Ho-W), and (d)
Kalman linear filter. The first presentation is based on a robust linear model (RLM) [6]. Like other
regression methods, its aim to model the dependent variables (1) by independent variables
ππ = π·ππππ + π·π πππ + β― + π·π πππ + πΊπ , (2)
where i = 1, ..., n, π₯π β is energy consumption of i-th consumer, π½0 , π½1 , β¦ π½π - regression coefficients,
π’π1,..., π’ππ - binary variables, ππ is a white noise. Extensions for regression model (2) are generalized
additive models (GAM) [8]
π³
π¬(ππ) = π·π + β π·π ππ (πππ ), (3)
π=π
where ππ are B-splines [8,14], πΏ β rank of regression model.
314
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
2.2. Cluster analysis for energy consumption in smart grids
Utility services have been differencing their consumers into industrial, commercial and
residential sectors based on some internal information on them. It may be nominal demand, the whole
consumption, etc. This empirical approach could be applied to define the sets of customers load
profiles and each user will be assigned to one of these profiles. The great problem to do this is, firstly,
that the consumption data of customers, who have installed smart meters, are now accessible. So , they
could change their profile dynamically. Secondly, the time period of measurement is not restricted and
usage information for some successive years is available. Thus, as the smart meters data are
continuously reported, they can have possible applications of real-time operation and control of energy
systems. All these factors stimulate the need to develop new clustering methods to characterize energy
consumption.
For classification consumers into groups (clusters), we used the centroid based clustering
method K-means [4,5,7]. The advantage of the K-means method is based on carefully seeding of
initial centroids, which improves the speed and accuracy of clustering [10]. Before applying the K-
means algorithm the optimal number of clusters k must be determined. For each representation of a
data set, we have determined the optimal number of clusters to k using the internal validation rate
Davies-Bouldin index [10,13]. It works as follows. Let π(π₯) denote the shortest Euclidean distance
from a data point x to the nearest centroid we have already chosen. Choose an initial centroid K1
uniformly at random from X. Choose the next center πΎπ = π₯Μ β π, selecting with probability π(π₯Μ)2 /
βπ₯βπ π(π₯)2 . Repeat previous step until we have chosen a total of K centers. Each object from data set
is connected with a centroid that is closest to it. New centroids are then calculated. Last two steps are
repeated until classification to clusters no longer changes. In each iteration we have automatically
determined the optimal number of clusters.
2.3. Energy consumption time series forecasting
We used three methods to improve forecasting energy consumption time series: Support
Vector Regression (SVR) [6], a method based on a combination of STL decomposition, Holt-Winters
exponential smoothing and ARIMA model [9]. Seasonal decomposition of time series based on
LOESS regression is a method, which decomposes seasonal time series into three parts: trend,
seasonal component and remainder (noise). For the final three time series any of the forecast methods
is used separately, in our case either Holt- Winters exponential smoothing or ARIMA model [12].
Random Forest (RF) algorithm is suitable for classification and regression [15]. The method constructs
the large number of decision trees at training time. Its output is the class that is the mode of the classes
(classification) or mean prediction (regression) of the individual trees. Gradient Boosting Machine
(GBM) is an efficient and scalable implementation of gradient boosting framework by Friedman [15].
The GBM was first proposed for classification problems. Its basic principle is that several simple
models, called weak learning models, to be merged into one iterative scheme for the selection of
parameters with the aim of obtaining the so-called strong learning model, i.e. models with better
prediction accuracy. Thus, the GBM algorithm iteratively adds a new decision tree (i.e. βweak
learnerβ) at each step, which best reduces the loss function. Specifically, in a regression model, the
algorithm starts with model initialization, which is typically a decision tree minimizing the loss
function (RMSE), and then at each step, a new decision tree is adjusted to the current residual and
added to the previous model to update the residuals. The algorithm continues to run until the
maximum number of iterations is reached or the specified precision is reached. It means that at each
new step, the decision trees added to the model in the previous steps are fixed. Thus, the model can be
improved in those parts of it where it still does not assess the residuals.
Bagging (Bagg) predictors generate multiple versions of predictors and use them for
determination an aggregated predictor, so the aggregation is an average of all predictors [16]. The
bagging method gives substantial gains in accuracy, but the vital element is the instability of the
prediction method. In the case that perturbing the learning set has significant influence on the
constructed predictor, the bagging can improve accuracy. The accuracy of the forecast of electricity
consumption was measured by MAPE (Mean Absolute Percentage Error). MAPE is defined as
follows:
315
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
π
π |ππ β π
Μ
|
π΄π¨π·π¬ = β πππ% (4)
π ππ
π=π
where π₯π is actual consumption, π₯Μ
β load forecast.
3. Computer experiments for energy consumption forecasting
We performed advanced computer experiments to evaluate the profit of using clustering
procedures on four time series representation methods for one day ahead forecast. Table 1 shows
average daily MAPE forecast errors of 6 forecasting methods. Each forecasting method was evaluated
on 5 datasets; 4 datasets are clustered with different representation methods (Kalman, Ho-W, GAM,
RLM) and 1 dataset is aggregated electric load consumption (Aggr). The following conclusions can be
derived from Table 1. Optimized clustering of consumers significantly improves accuracy of forecast
with forecasting methods SVR, Bagging, GBM. Despite this, clustering with LOESS+ARIMA, RF, R-
Tree does not really improve accuracy of forecast. Robust representation methods of time series
Kalman, GAM and RLM performed best among all representations, while Ho-W was the worst in
most of the cases. The best result of all cases achieved by GBM algorithm with optimized clustering
using GAM representation with mean daily MAPE error under 3,44%.
Table 1. Mape(%)-error forecasting methods for aggregated load consumption. repres.: model-
based presentations of consumption time series. meth.: forecasting methods applied with clustering
METH.\REPRES Kalman Ho-W GAM RLM Aggr
LOESS+ARIMA 4.873 4.947 4.423 4.674 4.854
SVR 4.073 4.072 4.42 4.216 4.621
Bagging 3.438 3.475 4.23 3.34 4.34
GBM 3.78 4.036 3.44 4.21 4.46
R-Forest 4.479 4.476 4.36 4.62 4.72
R-Tree 4.42 4.476 4.33 4.26 4.63
3. Conclusion
Improving the accuracy of electricity consumption forecasts is an essential direction in the
development of intelligent energy systems. Machine learning methods such as cluster analysis,
boosting and others were used to implement this task. The main purpose of this work was to show that
the application of the consumer clustering procedure to the representation of time series of energy
consumption can improve the accuracy of their energy consumption forecasts. Robust linear model,
generalized additive model, exponential smoothing and Kalman linear filter were used as such
representations. In this paper we applied a modified K-means++ algorithm to more accurately select
centroids and the Davis-Boldin index to evaluate clustering results. Numerical experiments have
shown that the methods of forecasting such as LOESS+ARIMA, SVR, RF, Bagging considered in the
paper are more effective for improving forecast accuracy if used together with clustering. Prediction
methods performed the best reliable representations of RLM, GAM, and Kalman filter. The lowest
prediction error is obtained by GBM algorithm with the GAM presentation. Among the perspective
applications of clustering for smart grids are benefits for individual tariffs design, compilation smart
demand programs, improvement of load forecast, classifying new or non-metered consumers and other
tasks.
316
Proceedings of the VIII International Conference "Distributed Computing and Grid-technologies in Science and
Education" (GRID 2018), Dubna, Moscow region, Russia, September 10 - 14, 2018
References
[1] Haben S., Singleton C., Grindrod P., Analysis and Clustering of Residential Customers Energy
Behavioral Demand Using Smart Meter Data // IEEE Transactions on Smart Grid. 2015. V.PP. no.99.
P.1-9.
[2] Chicco G., Napoli R. and Piglione F., Comparisons Among Clustering Techniques for Electricity
Customer Classiffcation // IEEE Trans. Power Sys., 2013, vol. 21. P.933-940.
[3] Gelling C. W., The Smart Grid: Enabling energy effciency and demand response. The Fairmont
Press Inc., 2009.
[4] Aghabozorgi S., Shirkhorshidi A., et al, Time-series clustering: A decade review, Information
Systems, 2015, vol. 53, P. 16-38.
[5] Shahzadeh A., Khosravi A., Nahavandi S., Improving load forecast accuracy by clustering
consumers using smart meter data, International Joint Conference on Neural Networks (IJCNN), 2015,
P. 1-7.
[6] Andersen A., Modern Methods for Robust Regression. SAGE Publications, Inc, 2008.
[7] Wijaya T.K., Vasirani M. et al, Cluster-based aggregate forecasting for residential electricity
demand using smart meter data, in 2015 IEEE International Conference on Big Data. IEEE, oct 2015,
pp. 879-887.
[8] Wood S., Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC, 2006.
[9] Hyndman R.J., Koehler A.B., Snyder R.D. and Grose S., A state space framework for automatic
forecasting using exponential smoothing methods, International Journal of Forecasting, 2002, vol. 18,
no. 3, P. 439-454.
[10] Arthur D. and Vassilvitskii S., K-means++: The Advantages of Careful Seeding, in SODA '07
Proceedings of the eighteenth annual ACM-SIAM symposium on discrete algorithms, 2007, P. 1027-
1035.
[11] Hong W.C., Intelligent Energy Demand Forecasting, London: Springer Verlag, 2013.
[12] Taylor J.W., Short-term electricity demand forecasting using double seasonal exponential
smoothing, Journal of Operational Research Society, 2003, vol. 54, P. 799-805.
[13]Shchetinin E.Yu., Cluster-based energy consumption forecasting in smart grids, Springer
Communications in Computer and Information Science (CCIS), 919, 46-656. Springer, Berlin, 2018.
[14] Shchetinin Eu.Yu., Lyubin P.G., Fast two-dimensional smoothing with discrete cosine transform,
Springer Communications in Computer and Information Science (CCIS), Springer: Berlin, 2016, 678,
P. 646-656.
[15] Breiman L., Random forests, Machine learning, 2001, vol. 45(1), P.5β32.
[16] Breiman L., Bagging predictors, Machine learning, 1996, vol. 24(2), P. 123β140.
317