DOI: 10.3303/CET2188124 
 

 

 
 
 

 
 
 

 
 
 

 
 

 
 
 

 
 
 

 
 

 

 
 
 

 
 

 
 
 

 
 
 

 
 
 

 
 
 

 

Paper Received: 29 April 2021; Revised: 29 September 2021; Accepted: 10 October 2021 
Please cite this article as: Kirillova E.A., Lazarev A.I., 2021, Innovative Energy Management System for Mobile Processors Power Consumption 
with Integration of Predictive Neural Network Models, Chemical Engineering Transactions, 88, 745-750  DOI:10.3303/CET2188124 

CHEMICAL ENGINEERING TRANSACTIONS 

VOL. 88, 2021 

A publication of 

The Italian Association 
of Chemical Engineering 
Online at www.cetjournal.it 

Guest Editors: Petar S. Varbanov, Yee Van Fan, Jiří J. Klemeš

Copyright © 2021, AIDIC Servizi S.r.l. 

ISBN 978-88-95608-86-0; ISSN 2283-9216 

Innovative Energy Management System for Mobile 
Processors Power Consumption with Integration of Predictive 

Neural Network Models 
Elena A. Kirillova*, Alexey I. Lazarev 
Department of Information Technology in Economics and Management, branch of the National Research University 
Moscow Power Engineering Institute” in Smolensk, Energy passage 1, 214013 Smolensk, Russia 
no@sbmpei.ru 

At the present stage of information technologies development in the field of portable devices implementation 
one of the perspective direction is the development of software and hardware to optimise the consumption of 
energy resources in mobile processors. Modern solutions are aimed at continuous improvement of methods to 
reduce power consumption along with increasing the performance of devices, but the problem of limited time 
interval of portable devices active use is still relevant. Existing solutions are aimed at using various energy-
dependent methods in the kernel configuration – I/O scheduler governors, TCP overload algorithms, as well 
as additional entropy distribution systems and memory release algorithms. The solutions considered involve a 
system of aggressive behaviour based on the use of user-driven methods to terminate energy demanding 
processes in the system. The purpose of this study is to overcome the problems under consideration by 
developing a cross-platform algorithmic approach, which is based on the tracking of the energy consumption 
processes in the kernel, taking into account system calls and background activity in the system. A distinctive 
feature of the proposed solution is the use of a neural network training sample for the processes of tracking 
user behaviour, which affects the reduction of the CPU load by completing side processes and increasing the 
time interval of battery performance. To implement the project, root access to the system was also used, 
assuming full-function access to the system kernel. Another feature of the implemented algorithm is 
backwards compatibility of the work with mobile processors, which allows to organise work on mobile devices. 

1. Introduction

The current development of information technologies is primarily aimed at the market for the production of 
innovative products in the field of mobile architecture. Manufacturers of software and hardware solutions are 
currently improving algorithms and methods to improve system performance in parallel with reducing power 
consumption. This area also affects the electric power and chemical industries due to the automation of 
individual sub-processes of the operation of the entire system, in particular, when using uninterruptible power 
supply systems to target installations. Among other things, Rego et al. (2019) describes a methodology for 
using neuro-fuzzy forecasting of biogas in the chemical industry, and Monteiro et al. (2018) describes the 
possibility of neural network data processing to detect measurement errors when preparing experimental 
samples. The considered methods and algorithms, integrated into mobile devices, offer the user a flexible 
system for ensuring efficient operation over a long period of time, up to the use of artificial intelligence for 
training and optimising energy consumption algorithms. However, the problem of active use of the device in a 
short time interval remains relevant. 
The solution to this problem from third-party developers is the use of various energy-dependent methods, 
among which there are governors, I/O schedulers, TCP overload algorithms, entropy distribution systems and 
volatile memory release. Examples of control system implementations in the Android operating system can be 
software solutions from the developers of ElementalX, Franco Kernel. Also important is the solution to the 
problems of centralised CPU load balancing described by Zhao et al. (2020). All the considered indicators are 
controlled at the kernel level and allow you to significantly reduce the coefficient of active and passive energy 

745



hourly consumption, while the user may face obvious problems of insufficient memory allocation, excessive 
increase or decrease in efficiency due to the choice of an incorrect governor. The publication of Harini and 
Ravikumar (2020) demonstrates the dependence of dynamic scaling of the voltage and frequency of the CPU 
in order to increase efficiency through parallel processing. Rothman (2021) describes the main provisions of 
implementing neuro-fuzzy systems in software solutions. The value prediction model uses as the main 
process a bidirectional recurrent neural network, described by Nathwani (2020) for solving signature 
recognition problems, and allows determining the dependence between parameters based on a training 
sample. 
Although the solution of the described problem is currently being paid quite close attention by Russian and 
foreign researchers, all existing methods have a common drawback, namely, they assume the user 
completion of energy-demanding processes in the system, which is not always convenient and possible. The 
algorithms used in the standard representation to reduce CPU and GPU load are characterised by a critical 
change in the operating mode due to the accurate determination of the values of system parameters in real 
time, which in turn does not always affect the optimal operation of a particular device. This determined the 
research objective to design a cross-platform system based on the analysis and management of system calls 
and background activity of the device without the direct involvement. From a technical point of view an 
essential feature of the proposed solution is the modular architecture for tracking system load statistics. The 
system is based on an algorithm that allows using artificial neural networks and fuzzy logic to determine the 
critical points of increasing the processor frequency, adjusting the kernel parameters, and ensuring stable 
operation of the processor with minimal energy consumption, which determines the theoretical novelty of the 
proposed method. The practical significance is determined by the fact of tracking user and system processes 
with the possibility to reduce energy consumption. 

2. Algorithmic implementation of the system

The main requirements highlighted for the algorithm to be implemented are multiplatform and the ability to 
work with most ARM family processors. A secondary requirement for the implemented algorithm is to have 
root access to the file system due to the need to handle kernel parameters and system processes, described 
in the paper on kernel optimisation by Yuan et al. (2018). An alternative important feature of the implemented 
algorithm is support for working with custom kernels and governors on the Android OS platform, due to the 
direct dependence of CPU regulators on energy efficiency (Hoffman et al., 2020). 
The algorithm for identifying energy-dependent parameters is based on determining the type of available 
options and then presenting them as factors processed by the neural network. To implement a modular 
architecture for tracking background processes, a hierarchical algorithm was designed, as illustrated in Figure 
1. The presented algorithm is processed by parallel processing of the following modules:
 Module “define available kernel functions” is used to initially identify the system parameters supported by

the kernel. These parameters include CPU and GPU controls, stun and boost options, as well as
minimum and maximum allowable CPU frequencies;

 Module “track usage statistics” allows you to accumulate information about user activity, the most energy-
significant processes and determine the level of CPU load;

 Module “predictive identification” module is used to predict the values of various factors for subsequent
decision-making;

 Module “decision-making” optimises the system and then compares the results;
 Module “optimising the operation of the governor” assumes the possibility of changing the system

governor to force a decrease or increase in the processor frequency.

Figure 1: The algorithm of the implemented system for reducing energy consumption 

746



2.1 Statistical data accumulation module 

The data accumulation module intelligently determines the moments of user activity. The main activation 
function of this module is the peak moments of unloading and loading processes and the analysis of the 
current load level. This module involves converting the parameters to a string form, which are then written to 
the corresponding database table. The database is represented by a PostgreSQL server deployment based 
on the construction principles described by Schönig (2020). As can be seen from Figure 2, each parameter 
group was divided into separate tables with corresponding relationships. The adaptation of the structure under 
consideration made it possible to record the original and predicted patterns in a single hierarchy. 

Figure 2: ER-diagram 

2.2 Predicting available parameters 

The predictive parameter processing module in this software uses the prediction of system calls that have a 
direct impact on the excessive increase in the discharge rate. As you can see from Figure 3, the non-
optimised operation of the kernel at certain points in time causes the discharge rate to increase to 30 %/h. 
This aspect of the work mainly affects the use of the processor governor, which increases the CPU frequency 
to the maximum, as well as the dependencies of the memory options table in the structure of the ER diagram.  

Figure 3: Analysis of abnormal operation of the energy consumption system 

According to research by Yan et al. (2019), the methodology for measuring energy consumption is divided into 
several detailed procedures, analytically described by the combination of consumption of various factors in Eq 
(1). The Esmart indicator describing the power consumption based on the processor is reflected by the Eq (2). 

(1) 

were Etotal – total power consumption, W; Esmart – CPU-based power consumption, W; EBS – mobile base 
station power consumption, W; Ewireline – mobile wired power consumption, W; EDC – mobile data centre power 
consumption, W. 

(2) 

747



were Esmart – total power consumption, W; Dsig – signal duration, s; PS_CPU – processor power consumption, W; 
PS_4G – 4G module power consumption, W. 
Since the parameters under consideration have a dynamic representation of the float type, their prediction is 
carried out using a bidirectional recurrent neural network (BRNN), as described in the publication Bohan and 
Yun (2019) of time series forecasting. An essential feature of the application of this neural network is the 
ability to predict input data in a bidirectional time interval due to the implementation of recurrent directed 
neurons that work in the opposite direction. Among other things, Long-Short Term Memory (LTSM) belongs to 
the category of a neural network with a long-term memory type, which allows you to use the dependencies 
identified in a variable time interval in order to record notable features, described by Goldberg (2017). The 
standard implementation of LTSM is based on the use of the following equations: 

(3) 

(4) 

(5) 

(6) 

(7) 

(8) 

were xt – input vector; ht, ht-1 – time output vector «t», referring to the value of t-1; it – output activation vector; 
Wix, Wih, b – weights of the representation of vectors for each step. 
The software implementation of the considered neural network is based on the use of the machine learning 
library PyTorch Pointer (2019). The validity of the choice of the PyTorch library is due to the advantage of the 
software based on the Python language, which allows you to run program code on multi-platform operating 
systems-based on the Linux kernel, OSX, as well as the Microsoft Windows OS family. The input vector (xi) 
processes the input data - the CPU load, as well as other dynamic parameters of the kernel using the hidden 
states of the forward and reverse inputs, in this connection, the activation vector of the hidden layer takes the 
form Eq (9 – 10): 

(9) 

(10) 

were W – the weight of the argument in forward and backward propagation; h – the weight of the hidden layer 
for processing the value in the time intervals t-1, t+1; b – offset vector. 
So, the input data for this neural network are floating values, for example, the CPU is represented as “1,267; 
1,363; 1,925”, while the output data has an identical form of predicted values. To operate a bidirectional ANN, 
the layers.Bidirectional function is added with the appropriate parameters, with a function value of zero for the 
reversed layer. The processing of the layer weight on a single interval with the subsequent use of the reverse 
gru function is noteworthy. Testing of the implemented software component was carried out using an Intel 
Core i7-10510U processor, as well as a GeForce GTX 1650 Max-Q design video card compatible with CUDA 
technology. 
As a result of the operation of a single instance of the output neural network, the output vector represents 
floating values, which are represented by inputs in the fuzzy logic apparatus. Table 1 shows the results of 
neural network training, according to which the data “accuracy” and “losses” increase and decrease in the 
process of neural network training, which indicates successful training. 

Table 1: Neural network training results 

Patterns, n 50 100 200 400 600 800 1000 
Loss, % 1,743 1,562 1,424 1,125 0,984 0,812 0,772 
Accuracy, % 1,934 1,836 1,743 1,472 1,293 0,946 0,852 

748



2.3 Decision-making based on the fuzzy logic apparatus 

The fuzzy logic unit processes the output values of the neural network for appropriate decision – making on 
the termination of individual processes or changing the operating mode of the governor. Fuzzy rule 
interpolation is based on the Sugeno inference model, an improvement of which is proposed in the publication 
of the authors Zhang and Shen (2019). The software implementation of this module is based on the 
FuzzyWuzzy library, which allows the use of floating value character comparison functions for subsequent 
decision-making. Processing the output values of the neural network vectors allowed us to obtain the following 
results when calling the “fuzz.Ratio” function: 

cpu_state_current = [“806, 650, 1404, 520, 1300, 949, 2002, 520”] 

cpu_state_predictive = [“445, 650, 1508, 624, 1742, 949, 1222, 520”] 

fuzz.ratio(cpu_state_current, cpu_state_predictive) 

IF (DISCARGE_RATE > 10.0) THEN CALL MODULE_STATE; 

IF (E_TOTAL > 500) THEN CALL MODULE_ANALYZE; 

IF (DISCARGE_RATE > 10.0) AND (TIME > 3) THEN CALL MODULE_EOPTIMAL; 

IF (DISCARGE_RATE > 15.0) AND (TIME > 7) THEN CALL MODULE_ESAVING; 
The version of the rules presented above also allows you to separately use the modules No. 4-5 in Figure 1, 
that is, it can be used both to change the main parameters of the kernel and to make changes to the operating 
mode of the governor. Testing of the operability of this set of rules was performed both on a personal 
computer with Microsoft Windows OS, and using the Android OS operating system with Python modules 
installed. As a result of the operation of this module, it can be concluded that the joint use of a neural network 
instance together with the fuzzy logic module allows you to make decisions on adjusting the operation of the 
device in order to reduce power consumption. 

2.4 Analysis of test results 

As a result of the development of the software module, the impact of the implemented software was tested on 
the Samsung Galaxy Note 10 (N970F) device. As can be seen from Figure 4, the integration of the developed 
algorithm allowed to reduce the process of energy consumption by eliminating the moments of excessive 
increase in the discharge rate parameter. 

Figure 4: Testing the performance of the implemented software 

Figure 4 also shows the user's activity period, which varies from 10 h (the original graph) to several days (the 
resulting graph). Also important is the process of reducing the recharge cycles of the Li-ion battery, which 
affects the overall battery life. In the future, it is planned to conduct testing on other types of processors, 
finalize the proposed software solution for a wider range of consumers and register such a software package. 

3. Conclusion

Currently, it is quite acute the problem of insufficient optimization of the active power consumption of mobile 
processors was highlighted. The methods proposed by Russian and foreign scientists to overcome it have 
significant drawback related to the need for active user interaction with the hardware and software system. In 
this regard, the authors proposed the development of a cross-platform system for optimising power 
consumption due to analyses of dynamic and static factors was carried out. The analysis of active energy 
consumption showed that the peak values of the energy consumption indicator per hour reach the level of 30 
%/h. During the analyses of devices operation also the main energy-dependent factors of the CPU, GPU, as 
well as application factors that are worked with using the PostgreSQL database were highlighted. The article 
presents an algorithm that optimizes the power consumption of mobile processors. 
The novelty of the proposed approach is a software solution for managing the operation of mobile processors 
to optimize energy consumption without active user intervention, which was implemented as follows. At the 

749



stage of predicting the available parameters, a bidirectional recurrent neural network was integrated into the 
implemented system, which made it possible to use the trained patterns to predict the main parameters of the 
system – active and passive energy consumption, processor frequencies to determine and reduce critical 
moments of electricity consumption. To reduce energy consumption during operation, a method of neural 
processing based on previously recorded patterns was introduced. The implementation of the module of the 
fuzzy logic and decision-making apparatus is based on the use of specialised algorithms and subsequent 
decision-making based on the conditional comparison rules given in the article. The software implementation 
of the modules under consideration was carried out using the Python libraries – PyTorch, FuzzyWuzzy 
described by Vishnu (2018). 
Confirmation of the effectiveness of the developed algorithm is reflected in the graphs of the analytical 
comparison of energy consumption, where the indicator of active energy consumption was reduced to 2%/h, 
resulting in the conclusion that the algorithm can be implemented in systems with an ARM processor. 
The developed approach reveals wide opportunities for the restructuring of energy management systems for 
mobile processes. A promising increase in the efficiency of ARM systems, along with a decrease in energy 
consumption, is possible due to the development of a universal behaviour model based on multiple models of 
daily operation of mobile devices. The development of a universal neural network model will optimise the 
process of mobile penetration testing, where energy consumption has a significant impact on the environment 
of the chroot container. The development in the IoT segment will allow us to organise a more efficient system 
of autonomous object management in enterprises of various sizes. 

References 

Bohan H., Yun B., 2019. Traffic Flow Prediction Based on BRNN, 9th International Conference on Electronics 
Information and Emergency Communication (ICEIEC), Beijing, China, 12-14 July, 320-323. 

Goldberg Y., 2017. Neural network methods in natural language processing, Morgan & Claypool Publishers, 
San Rafael, CA, USA. 

Harini S., Ravikumar A., 2020. Effect of Parallel Workload on Dynamic Voltage Frequency Scaling for Dark 
Silicon Ameliorating, International Conference on Smart Electronics and Communication (ICOSEC), 
Trichy, India, 10-12 September, 1012-1017. 

Hoffman A., Pathania A., Kindt P.H., Chakraborty S., Mitra T, 2020. BrezeFlow: Unified debugger for android 
CPU power governors and schedulers on edge devices, 57th ACM/IEEE Design Automation Conference 
(DAC), San Francisco, USA, 5-9 December, 1-6. 

Monteiro F., Valim I., De Siqueira R., Moura F., Grillo A., Santos B., 2018. Application of Artificial Neural 
Networks for Identification of Catalysts Used in Thermogravimetry Lignocellulosic Biomass, Chemical 
Engineering Transactions, 65, 529-534. 

Nathwani C, 2020. Online Signature Verification Using Bidirectional Recurrent Neural Network, 4th 
International Conference on Intelligent Computing and Control Systems (ICICCS), Madurai, India, 13-15 
May, 1076-1078. 

Pointer I., 2019. Programming pytorch for deep learning: Creating and deploying deep learning applications, 
O’Reilly Media, Farnham, UK. 

Rego A., Leite S., Leite B., Grillo A.V., Santos B.F., 2019. Artificial Neural Network Modelling for Biogas 
Production in Biodigesters. Chemical Engineering Transactions, 74, 25-30. 

Rothman D., 2021. Transformers for natural language processing: build innovative deep neural network 
architectures for Nlp with Python, PyTorch, TensorFlow, Bert, RoBERTa, and more, Packt Publishing, 
Birmingham, England, UK. 

Schönig H., 2020. Mastering PostgreSQL 13: Build, administer, and maintain database applications efficiently 
with PostgreSQL 13, 4th ed., Packt Publishing, Birmingham, England, UK. 

Vishnu S., 2018. Deep Learning with PyTorch: A practical approach to building neural network models using 
PyTorch, Packt Publishing, Birmingham, England, UK. 

Yan M., Chan C.A., Gygax A.F., Yan J., Campbell L., Nirmalathas A., Leckie C., 2019. Modeling the total 
energy consumption of mobile network services and applications, Energies, 12 (1), en12010184. 

Yuan P., Guo Y., Chen X., Mei H., 2018. Device-Specific Linux Kernel Optimization for Android Smartphones, 
6th International Conference on Mobile Cloud Computing, Services, and Engineering (MobileCloud), 
Bamberg, Germany, 26-29 March, 65-72. 

Zhang P., Shen Q., 2019. A Novel Framework of Fuzzy Rule Interpolation for Takagi-Sugeno-Kang Inference 
Systems, International Conference on Fuzzy Systems, (FUZZ-IEEE), New Orleans, USA, 23-26 June, 1-6. 

Zhao S., Hao C.-L., Zhai J., Li M.-S., 2020. S-Bridge: CPU Load Balancing Agent for Performance Asymmetric 
Multicore Processors. Ruan Jian Xue Bao/Journal of Software, 31 (9), 2965-2979. 

750