Machine Learning 2019: Smart system for forecasting energy power consumption - Aliya Amirzhanova - University Heidelberg, Germany

Forecasting plays an important role in many fields such as data science, economics, finance, supply chain, marketing, weather and nature conditions etc. Planning and preparing for the future is a most critical and important part in any organization. Forecasting itself is the process of predicting future trends, values based on a collection of observations made sequentially through equally spaced periods of time.

Today’s one of the big topics on the table is electricity consumption. In evidence, throughout the EU as one of the steps of energy policy goals by 2020 are smart electricity solutions to achieve objectives in more intelligent control over electricity supply and consumption [1]. This can be done by smart electricity forecasting system, which is based on individual household consumption, allowing not only saving people’s budget and encouraging them to use less electricity, but also might impact on a higher, state level of electricity saving. Thus, to achieve leveraging metering solutions, this work proposes to implement Long Short-Term Memory network (LSTM) and Gated Recurrent Unit network (GRU), as a solution for the smart control systems. These NN can outperform classical statistical approaches and show an ability to predict future consumption very well. Moreover, a novel hybrid model of using CNN+GRU technique, which previously was only used in a classification task [2, 3], has been performed in this work, and promising results, which even compete with baseline approaches, are obtained.

Creating powerful vitality asset the executives systems in the brilliant network is trying because of the elements on both the interest and gracefully sides encountering various changes. In this paper, we address the issue of evaluating vulnerabilities on the vitality request side. In particular, we initially create approaches utilizing factual displaying investigation to infer a measurable dispersion of vitality use. We at that point use a few AI based methodologies, for example, the Support Vector Machines (SVM) and neural systems to do precise estimating on vitality use. We perform broad analyses of our proposed approaches utilizing a certifiable meter perusing informational index. Our test information shows that the measurable dissemination of meter perusing information can be to a great extent approximated with a Gaussian dispersion and the two SVM-based AI ways to deal with accomplish a high precision of determining vitality use. Expansions to other savvy network applications (e.g., guaging vitality age, deciding ideal interest reaction, and abnormality identification of malevolent vitality use) are examined also. To build up a measurable model of vitality use, we create two non-parametric test based ways to deal with infer the factual dispersion of vitality use dependent on the previously mentioned true meter understanding information. We utilize a non-parametric test to complete the investigation of the vitality use information. For a lot of one-dimensional information, basic non-parametric test approaches incorporate the Shapiro-Wilk test [31] and the Kolmogorov-Smirnov (K-S) test [12]. It is important that on the grounds that the K-S test requests the preknowledge of the circulation of the example information, the test outcome won't be solid if the populace's Cumulative Distribution Function (CDF) is evaluated from the example information. It is significant that the foreordained CDF of the meter information isn't known, so we consider the Shapiro-Wilk test to test the dissemination of the example information. We likewise utilize another non-parametric test approach, which is additionally called Quantile-Quantile (Q-Q) plot ordinariness test, to affirm the appropriation of meter perusing information [36]. On the plot, when two informational collections are indistinguishably disseminated, the Q-Q plot will be demonstrated a line. At that point, we realize that the more prominent the takeoff from the reference line, the more noteworthy the possibility that the two informational indexes are drawn with various conveyances.

As a rule, the SVM is one of the famous techniques to productively group information and to construct a classifier, which can be additionally used to do guaging. In SVM, the information and related highlights can be treated as a point and vectors in multi-dimensional space. The fundamental rule of a standard SVM is to discover a hyperplane, which could separate the focuses into various spaces. Thusly, we can group information into various classifications [27]. So as to limit the arrangement mistake, the best possible hyperplane should be resolved. The least squares SVM that is additionally meant as LS-SVM is an upgraded SVM [33]. In a LS-SVM, there are two significant upgrades in examination with the standard SVM. To start with, the imbalance imperatives are subbed by equity requirements. Second, the squared misfortune work is utilized in the target work [34]. In our trial, we utilize the spiral premise work as the piece work in LS-SVM because of its wide use. In this segment, we present the exhibition assessment results. We initially present the trial arrangement and afterward present the consequences of measurable displaying and vitality utilization In this segment, we present the presentation assessment results. We initially present the trial arrangement and afterward present the aftereffects of factual displaying and vitality utilization.

Electric utilization determining utilizing savvy meter dataset is one of the angles in which AI approach is profoundly applied. Determining top interest and electric machine utilization requires nitty gritty examination of keen meter information through grouping and bunching strategies. Determining of electrical apparatus and Peak request is fundamental activity and a noteworthy part in electric force framework arranging and advancement. Be that as it may, because of changeability of family unit utilization level interest and apparatus utilization request, profound and detail examination of clients' shrewd meter information is required so as to recognize basic traits and the wellspring of variety between the utilization level of machine, just as clients request. This paper centers around anticipating levels of electric machine utilization and pinnacle request with the way of life of private client's utilizing information acquired from Irish and Umass archive. Further on, clients way of life is broke down from the consequences of client top interest conjecture. Directed and unaided AI calculation called CLARA grouping, bolster vector machine (SVM) and counterfeit neural system are applied as so as to accomplish gauge the apparatus utilization level and pinnacle request. Mean electric machine utilization esteems are determined from day by day, week after week, month to month and all out utilization for every apparatus from 1 year shrewd information of 1 min time stretch for electric apparatus utilization anticipating of individual family units. For the clients' pinnacle request utilization, just mean of week by week utilization of totaled family units is figured together. The guaging of clients electric utilization utilizing SVM gives result of 99.6% exactness which is far superior to the past works in a similar field of study. The got outcome shows that the executed philosophies and calculations are applied at their best degree of execution.


Author(s): Aliya Amirzhanova

Abstract | PDF

Share This Article