Сross-domain adaptation and trusted quality assurance of intelligent energy management software
Intelligent Energy Management Software must ensure reliable operation across heterogeneous domains where data distributions and system environments frequently vary. This paper introduces a cross-domain adaptation and trusted quality assurance framework that combines transfer learning, adversarial do...
Saved in:
| Date: | 2026 |
|---|---|
| Main Authors: | , |
| Format: | Article |
| Language: | English |
| Published: |
Інститут проблем реєстрації інформації НАН України
2026
|
| Subjects: | |
| Online Access: | https://drsp.ipri.kiev.ua/article/view/363165 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Journal Title: | Data Recording, Storage & Processing |
| Download file: |  |
Institution
Data Recording, Storage & Processing| _version_ | 1868294457926877184 |
|---|---|
| author | Жихай, Ванг Верлань, А. А. |
| author_facet | Жихай, Ванг Верлань, А. А. |
| author_institution_txt_mv | [
{
"author": "Ванг Жихай",
"institution": "Guangzhou Xinhua University, National Technical University of Ukraine \"Ihor Sikorsky Kyiv Polytechnic Institute\""
},
{
"author": "А. А. Верлань",
"institution": "National Technical University of Ukraine \"Ihor Sikorsky Kyiv Polytechnic Institute\", Norwegian University of Science and Technology"
}
] |
| author_sort | Жихай, Ванг |
| baseUrl_str | http://drsp.ipri.kiev.ua/oai |
| collection | OJS |
| datestamp_date | 2026-06-17T17:50:04Z |
| description | Intelligent Energy Management Software must ensure reliable operation across heterogeneous domains where data distributions and system environments frequently vary. This paper introduces a cross-domain adaptation and trusted quality assurance framework that combines transfer learning, adversarial domain alignment, and calibration based reliability assessment, validated on NASA MDP, PROMISE, NAB, and UCI energy benchmarks. Compared with strong baselines such as Random Forest, Convolutional Neural Networks, and Gated Recurrent Units, the proposed method achieves consistent improvements, yielding absolute F1 score gains of 5 to 10 points on defect prediction (NASA MDP, PROMISE) and an increase of 8 points on anomaly detection (NAB, from 0.70 to 0.78), while reducing Expected Calibration Error to 0.032 (a 22 to 42 percent reduction relative to Bayesian CNN baselines) and Negative Log Likelihood to 0.18, thereby demonstrating that integrating cross-domain adaptation with rigorous quality assurance mechanisms significantly enhances both predictive performance and reliability in real world IEMS deployments. |
| doi_str_mv | 10.35681/1560-9189.2026.28.2.363165 |
| first_indexed | 2026-06-18T01:00:39Z |
| format | Article |
| fulltext |
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 79
DOI: https://doi.org/10.35681/1560-9189.2026.28.2.363165
UDC 004.8:004.056.5:621.311
Wang Zhihai1,2 , A. A. Verlan2,3
1Guangzhou Xinhua University
Guangzhou, China No. 19 Huamei Road, Longdong, Tianhe District, 510520
2National Technical University of Ukraine «Ihor Sikorsky Kyiv Polytechnic Institute»
37, Prospect Beresteiskyi, 03056 Kyiv, Ukraine
3Norwegian University of Science and Technology
Teknologiveien 22, 2815 Gjøvik, Norway
Сross-domain adaptation and trusted quality assurance
of intelligent energy management software
Intelligent Energy Management Software must ensure reliable operation
across heterogeneous domains where data distributions and system environ-
ments frequently vary. This paper introduces a cross-domain adaptation and
trusted quality assurance framework that combines transfer learning, adver-
sarial domain alignment, and calibration based reliability assessment, vali-
dated on NASA MDP, PROMISE, NAB, and UCI energy benchmarks. Com-
pared with strong baselines such as Random Forest, Convolutional Neural
Networks, and Gated Recurrent Units, the proposed method achieves con-
sistent improvements, yielding absolute F1 score gains of 5 to 10 points on
defect prediction (NASA MDP, PROMISE) and an increase of 8 points on
anomaly detection (NAB, from 0.70 to 0.78), while reducing Expected Cali-
bration Error to 0.032 (a 22 to 42 percent reduction relative to Bayesian CNN
baselines) and Negative Log Likelihood to 0.18, thereby demonstrating that
integrating cross-domain adaptation with rigorous quality assurance mecha-
nisms significantly enhances both predictive performance and reliability in
real world IEMS deployments.
Key words: Intelligent Energy Management Software (IEMS); Cross-Domain
Adaptation; Transfer Learning; Domain-Adversarial Training; Trustworthy
Artificial Intelligence; Software Quality Assurance; Calibration; Explaina-
bility.
1. Introduction
Intelligent Energy Management Software (IEMS) plays an increasingly important
role in modern smart grid systems, distributed energy infrastructures, and energy optimi-
zation platforms. With the rapid deployment of renewable energy resources, heteroge-
neous sensors, and intelligent control systems, IEMS must operate reliably across dy-
namic and heterogeneous environments.
© Wang Zhihai, A. A. Verlan
https://doi.org/10.35681/1560-9189.2026.28.2.363165
Wang Zhihai, A. A. Verlan
80
However, practical deployment environments often exhibit substantial differences
in data distributions, operational conditions, sensor configurations, and software infra-
structures. These cross-domain discrepancies may reduce model generalization ability
and negatively affect the reliability of intelligent decision-making systems.
Traditional quality assurance (QA) approaches are commonly designed for static
and homogeneous environments. Consequently, models trained in one deployment envi-
ronment may experience performance degradation when transferred to different opera-
tional domains. This limitation creates challenges for the reliable deployment of AI-
driven energy management systems.
Traditional quality assurance (QA) pipelines, however, were not built for this pur-
pose. They typically assume that what works in one facility will work in another, an as-
sumption that rarely holds. When a model trained in one plant is deployed to a different
site with unfamiliar sensors, distinct load behaviors, or mismatched software infrastruc-
tures, performance often degrades silently and unpredictably. The result is not merely an
academic inconvenience, but a practical vulnerability in systems that affect energy costs,
grid stability, and carbon emissions.
Recent studies have explored transfer learning and domain adaptation techniques to
address cross-domain discrepancies in intelligent systems [1, 2]. These approaches im-
prove model generalization ability across heterogeneous environments and reduce per-
formance degradation caused by distribution shifts [3].
In parallel, federated learning methods have been introduced to support distributed
collaborative model training while preserving data privacy [4, 5]. Such approaches are
particularly important for intelligent energy systems involving multiple distributed infra-
structures and heterogeneous operational environments.
In addition, trustworthy artificial intelligence has become an important research di-
rection in safety-critical software systems. Calibration assessment, robustness evaluation,
and explainability analysis are increasingly incorporated into quality assurance work-
flows to improve the reliability and interpretability of AI-driven systems [6–10].
Although substantial progress has been achieved in domain adaptation, federated
learning, and trustworthy AI, these approaches are often studied independently. Limited
research has investigated the integration of cross-domain adaptation mechanisms and for-
mal quality assurance frameworks within Intelligent Energy Management Software
(IEMS).
Therefore, a unified framework that simultaneously addresses cross-domain adap-
tation, reliability assessment, calibration, robustness, and explainability remains neces-
sary for practical IEMS deployment.
To address the above limitations, this study proposes a unified framework integra-
ting cross-domain adaptation and trusted quality assurance mechanisms for Intelligent
Energy Management Software (IEMS).
The main contributions of this study are summarized as follows:
1) а cross-domain adaptation framework integrating transfer learning and adversar-
ial domain alignment to improve model generalization across heterogeneous deployment
environments [1, 2];
2) а trusted QA layer incorporating calibration assessment, robustness evaluation,
and explainability analysis to improve the reliability and interpretability of intelligent en-
ergy management systems [6–10];
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 81
3) сomprehensive experimental validation on publicly available datasets, including
NASA MDP, PROMISE, NAB, and UCI Energy datasets, demonstrating improvements
in defect prediction, anomaly detection, and calibration performance [11–14].
The remainder of this paper is organized as follows. Section 2 reviews related work
in domain adaptation and trustworthy QA. Section 3 details the proposed framework.
Sections 4 and 5 present the experimental setup and results, respectively. Section 6 con-
cludes the paper and discusses the main implications and limitations of the proposed
framework.
2. Related Work
Research on cross-domain adaptation has gained substantial traction in computer
vision and natural language processing, where techniques such as adversarial training,
feature alignment, and transfer learning have demonstrated considerable effectiveness [1–
3]. Within software engineering, defect prediction has traditionally relied on supervised
learning confined to a single project, achieving only limited generalizability when applied
across projects [3, 11, 12]. To overcome these limitations, domain-adaptive approaches
— such as transfer component analysis and adversarial neural networks — have recently
been introduced and shown promising results in cross-project defect prediction [2, 13].
Parallel to these efforts, the field of trusted quality assurance has become increas-
ingly prominent. Calibration metrics, particularly Expected Calibration Error (ECE) and
Negative Log Likelihood (NLL), are now standard tools in evaluating trustworthy AI sys-
tems [6, 7]. Techniques including temperature scaling and Bayesian deep learning have
been proposed to enhance calibration performance [6]. In addition, robustness studies,
such as adversarial perturbation analyses, underscore the importance of QA frameworks
that extend beyond predictive accuracy alone [9]. Complementing these dimensions, ex-
plainability methods such as SHAP and LIME have been widely employed to improve
the transparency and interpretability of model predictions [10].
Despite these advances, relatively few studies have attempted to integrate cross-
domain adaptation with a formalized QA layer in the context of IEMS. This work ad-
dresses that gap by proposing a unified framework that combines domain adaptation with
calibration, robustness, and explainability mechanisms.
3. Methodology
This study proposes a two-layer architecture that integrates cross-domain adapta-
tion with a trusted quality assurance (QA) layer to enable reliable IEMS performance
across heterogeneous domains (Fig. 1).
To improve the reproducibility and interpretability of the proposed framework, the
overall cross-domain adaptation and trusted QA workflow algorithm is summarized as
follows.
Step 1. Source-domain and target-domain datasets are collected from heterogene-
ous intelligent energy management environments.
Step 2. Data preprocessing and feature normalization are performed to reduce in-
consistencies between heterogeneous data sources.
Step 3. Transfer learning is applied to initialize the target-domain model using
source-domain knowledge.
Wang Zhihai, A. A. Verlan
82
Step 4. Domain-adversarial alignment is employed to minimize feature distribution
discrepancies between heterogeneous domains.
Step 5. Federated aggregation is conducted to support distributed collaborative
learning without exposing raw data.
Step 6. The trusted QA layer evaluates calibration, robustness, and explainability
indicators.
Step 7. QA metrics, including Expected Calibration Error (ECE), Negative Log
Likelihood (NLL), and robustness margin, are calculated to assess model reliability.
Step 8. Final QA decisions are generated to support reliable deployment of Intelli-
gent Energy Management Software (IEMS).
Figure 1. Proposed cross-domain adaptation and QA framework
As illustrated in Fig. 1, the proposed framework processes heterogeneous data
through three sequential adaptation mechanisms: transfer learning, adversarial alignment,
and federated aggregation. Transfer learning leverages labeled source-domain data to in-
itialize target-domain models. Adversarial alignment minimizes domain-discriminative
features, thereby reducing distributional discrepancies across environments [1, 2]. Feder-
ated aggregation supports collaborative model updates across distributed utilities without
exposing raw data [4, 5]. The output of this stage consists of domain-adapted representa-
tions that are subsequently passed to the trusted QA layer.
The trusted QA layer applies three evaluation mechanisms: calibration, robustness
assessment, and explainability. Calibration is quantified using Expected Calibration Error
(ECE) and Negative Log Likelihood (NLL) to ensure that predicted probabilities reflect
true likelihoods [6, 8]. Robustness assessment evaluates model stability under input per-
turbations [9]. Explainability is achieved via SHAP and LIME, providing feature-level
interpretations of model outputs [8, 9]. These mechanisms function as quality checkpoints
within the deployment pipeline
3.1. Transfer Learning
Given a source domain Ds and a target domain Dt where Ps(x)≠Pt(x), we fine-tune
a pre-trained source model on limited target labels. Thus, we determine the Transfer
Learning Loss:
ℒ𝑇𝐿 = ℒsup(𝑓θ(𝑥𝑡), 𝑦𝑡) + 𝜆‖θ − θ𝑠‖2, (1)
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 83
where ℒ𝑇𝐿 denotes the transfer learning loss, ℒ𝑠𝑢𝑝 represents the supervised classification
loss, 𝑓θ(𝑥𝑡) denotes the prediction function parameterized by θ, yt represents the target-
domain labels, 𝜆 is the regularization coefficient, and ‖θ − θ𝑠‖2 constrains the target-do-
main parameters to remain close to the source-domain model parameters.
3.2. Adversarial Domain Adaptation
To minimize distributional shift, we apply adversarial feature alignment using a do-
main discriminator:
min
𝐹
max
𝐷
ℒ𝑎𝑑𝑣 = 𝔼𝑥𝑠∼𝐷𝑠
[log 𝐷 (𝐹(𝑥𝑠))] + 𝔼𝑥𝑡∼𝐷𝑡
[log( 1 − 𝐷(𝐹(𝑥𝑡)))], (2)
where F denotes the feature extractor, D represents the domain discriminator, xs∼Ds and
xt∼Dt denote source-domain and target-domain samples respectively, and Ladv represents
the adversarial domain adaptation loss.
3.3. Federated Learning
In distributed utility settings, federated learning aggregates local models without
raw data sharing:
θ(𝑡+1) = ∑
𝑛𝑘
𝑁
𝐾
𝑘=1 θ𝑘
(𝑡)
, 𝑁 = ∑ 𝑛𝑘
𝐾
𝑘=1 , (3)
where θ
(t+1) denotes the updated global model parameters; K represents the number of
participating clients; nk is the sample size of client k; N denotes the total number of samp-
les across all clients; θk
(t)
represents the local model parameters of client k at iteration t.
The Supervised Learning Objective is expressed as follows:
ℒsup = −
1
𝑛
∑ 𝑦𝑖
𝑛
𝑖=1 log 𝑝 (𝑦𝑖|𝑥𝑖), (4)
where ℒ𝑠𝑢𝑝 denotes the supervised learning loss; n represents the total number of training
samples; y
i
denotes the ground-truth label of sample i ;; p(y
i
|xi) represents the predicted
probability for the corresponding class given input xi.
3.4. Trusted QA Layer
Calibration, robustness and explainability are integrated into the following QA
pipeline:
1) Expected Calibration Error (ECE):
𝐸𝐶𝐸 = ∑
|𝐵𝑚|
𝑛
𝑀
𝑚=1 𝑎𝑐𝑐(𝐵𝑚) − 𝑐𝑜𝑛𝑓(𝐵𝑚), (5)
where M denotes the number of confidence bins; 𝐵𝑚 represents the set of samples be-
longing to the m-th confidence interval; n is the total number of samples; acc(𝐵𝑚) denotes
the prediction accuracy within bin 𝐵𝑚; conf(𝐵𝑚) represents the average confidence score
of predictions within the corresponding bin;
2) Negative Log Likelihood (NLL):
𝑁𝐿𝐿 = −
1
𝑛
∑ log 𝑝𝑦𝑖
(𝑛
𝑖=1 𝑥𝑖), (6)
Wang Zhihai, A. A. Verlan
84
where 𝑦𝑖 denotes the ground-truth label, 𝑝𝑦𝑖
(𝑥𝑖) represents the predicted probability for
the corresponding class, and N is the total number of evaluated samples;
3) Brier Score:
𝐵𝑟𝑖𝑒𝑟 =
1
𝑛
∑ ∑ (𝐶
𝑐=1 𝑝𝑖𝑐
𝑛
𝑖=1 − 1[𝑦𝑖 = 𝑐])2, (7)
where 𝑛 denotes the total number of samples; 𝐶 represents the number of classes; 𝑝𝑖𝑐 de-
notes the predicted probability that sample 𝑖 belongs to class 𝑐; 1[𝑦𝑖 = 𝑐] is an indicator
function that equals 1 when the true label of sample 𝑖 belongs to class 𝑐 and 0
otherwise;
4) Temperature Scaling:
𝑝 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑧/𝑇), 𝑇 > 0, (8)
where 𝑝 denotes the calibrated probability distribution; 𝑧 represents the model logits be-
fore calibration; 𝑍 is the temperature parameter used for probability calibration; soft-
max(·) denotes the softmax activation function. The condition 𝑇 > 0 ensures valid prob-
ability;
5) Entropy:
𝐻(𝑝) = − ∑ 𝑝𝑐
𝐶
𝑐=1 log 𝑝𝑐, (9)
where 𝐻(𝑝) denotes the entropy of the prediction probability distribution; p
c
represents
the predicted probability for class 𝑐; 𝑐 is the class index; 𝐶 denotes the total number of
output classes;
6) Robustness Margin:
Δ𝑅(𝜖) = 𝑅𝑐𝑙𝑒𝑎𝑛 − 𝑅𝑛𝑜𝑖𝑠𝑦(𝜖), (10)
where Δ𝑅(𝜖) denotes the robustness margin under perturbation level 𝜖; 𝑅𝑐𝑙𝑒𝑎𝑛 represents
the model performance on clean input data; 𝑅𝑛𝑜𝑖𝑠𝑦(𝜖) denotes the model performance
under perturbed or noisy conditions.
Together, these components constitute a unified pipeline that simultaneously en-
hances predictive transfer across domains while embedding formal QA checks to ensure
the trustworthiness of IEMS deployments.
4. Experimental Design
4.1. Datasets
We evaluate the proposed framework on four publicly available benchmarks: the
NASA MDP and PROMISE defect prediction datasets, the NAB time-series anomaly de-
tection benchmark, and the UCI energy dataset [11–14].
Table 1. Dataset overview
Dataset Domain Samples Features Labels
NASA MDP Software defects 7,797 38 Defect/No
PROMISE Software defects 10,500 45 Defect/No
NAB Sensor streams 58 Time-series Normal/Anomaly
UCI Energy Energy usage 19,735 28 Regression/Anomaly
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 85
Table 1 summarizes the datasets employed in our experiments. The NASA MDP
and PROMISE datasets support software defect prediction tasks. The NAB benchmark
enables evaluation of time-series anomaly detection. The UCI Energy dataset provides
real-world energy consumption measurements. This selection ensures comprehensive val-
idation across both software quality assurance and energy analytics domains. The quality
evaluation process is also conceptually aligned with ISO/IEC 25010 software quality as-
sessment principles [15].
4.2. Baselines
We compare RF [16], SVM, CNN, LSTM Autoencoder [17], GRU detector, and
Bayesian CNN.
Table 2. Baseline configurations
Model Type Key Parameters Reference
Random Forest Classical 200 trees, max depth = 10 [16]
SVM (RBF) Classical C = 1,0; γ = 0,1
CNN (baseline) Deep 3 conv layers, ReLU
LSTM Autoencoder Deep Hidden units = 128 [17]
GRU Detector Deep Hidden units = 128
Bayesian CNN Probabilistic MC-dropout, 20 samples
As listed in Table 2, the baseline models comprise classical machine learning algo-
rithms (Random Forest, SVM), deep learning architectures (CNN, LSTM, GRU), and a
probabilistic model (Bayesian CNN). This diverse selection provides a robust basis for
evaluating the proposed framework.
4.3. Metrics
We report the following classification metrics and calibration/robustness indicators:
1) Precision & Recall:
𝑃 =
𝑇𝑃
𝑇𝑃+𝐹𝑃
, 𝑅 =
𝑇𝑃
𝑇𝑃+𝐹𝑁
, (11)
where 𝑃 denotes Precision; 𝑅 denotes Recall; 𝑇𝑃 represents the number of true positives;
FP denotes false positives; 𝐹𝑁 represents false negatives;
2) F1-score:
𝐹1 =
2𝑃𝑅
𝑃+𝑅
, (12)
where F1 denotes the harmonic mean of Precision and Recall; 𝑃 represents Precision; and
R represents Recall.
4.4. Workflow
The overall experimental workflow is shown in Fig. 2.
Wang Zhihai, A. A. Verlan
86
Figure 2. The experimental workflow
As shown in Fig. 2, the experimental workflow proceeds sequentially through da-
taset selection, preprocessing, model training, quality assurance evaluation, and compar-
ative analysis of results.
5. Results and Discussion
As shown further, we report results across three task categories: defect prediction
(NASA MDP & PROMISE), anomaly detection (NAB), and calibration assessment (UCI
Energy).
5.1. Classification Performance (NASA MDP & PROMISE)
Table 3 presents accuracy, F1, and AUC. The proposed method yields absolute F1
gains of 5–10 points over baselines.
Table 3. Classification performance on NASA MDP & PROMISE
Model Accuracy (%) F1-score AUC
Random Forest 82,4 0,78 0,85
SVM (RBF) 80,1 0,74 0,82
CNN (baseline) 84,3 0,79 0,86
Proposed Method 89,7 0,84 0,91
Table 3 shows that the proposed method achieves the highest accuracy, F1-score,
and AUC among all baselines. In particular, it improves F1 by 5–10 percentage points
over Random Forest and CNN, confirming the benefits of cross-domain adaptation for
defect prediction.
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 87
As visualized in Figure 3, the proposed method consistently surpasses baseline clas-
sifiers across NASA MDP and PROMISE datasets. The improvements in both F1 and
AUC validate the robustness of cross-domain adaptation.
Figure 3. Classification performance (NASA MDP & PROMISE)
As visualized in Fig. 3, the proposed method consistently surpasses all baseline
classifiers across the NASA MDP and PROMISE datasets. The improvements in both F1
and AUC validate the robustness of the cross-domain adaptation approach.
5.2. Anomaly Detection (NAB)
Table 4 compares precision/recall/F1. The proposed method achieves F1 = 0,78 (vs.
0,70 for GRU).
Table 4. Anomaly detection results on NAB
Model Precision Recall F1-score
Isolation Forest 0,62 0,59 0,60
LSTM Autoencoder 0,70 0,65 0,67
GRU-based Detector 0,73 0,68 0,70
Proposed Method 0,81 0,75 0,78
As presented in Table 4, the proposed approach significantly outperforms conven-
tional methods in anomaly detection, reaching an F1-score of 0,78. Compared to GRU-
based models (F1 = 0,70), our method improves anomaly detection by 11 % relative gain.
Wang Zhihai, A. A. Verlan
88
Fig. 4 compares anomaly detection performance on NAB streams. The proposed
method demonstrates higher precision, recall, and F1-score than Isolation Forest, LSTM
Autoencoder, and GRU-based detector, proving its robustness under domain drift.
Figure 4. Anomaly detection results (NAB)
5.3. Calibration and Reliability (UCI)
Table 5 reports calibration metrics (lower is better). The proposed method achieves
ECE = 0,032 and NLL = 0,18.
Table 5. Calibration metrics on UCI
Model ECE NLL
Random Forest 0,062 0,29
CNN (baseline) 0,055 0,25
Bayesian CNN 0,041 0,22
Proposed Method 0,032 0,18
Table 5 highlights the calibration performance. The proposed method achieves the
lowest Expected Calibration Error (ECE = 0,032) and Negative Log-Likelihood (NLL =
0,18), demonstrating superior reliability for practical deployment in intelligent energy
management systems.
Fig. 5 presents calibration results. The proposed method yields the lowest ECE and
NLL values, confirming that its probability estimates align closely with true event fre-
quencies, which is critical for trustworthy predictions in energy management.
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 89
Figure 5. Calibration performance across models (UCI)
5.4. Discussion
The integration of cross-domain adaptation and trusted quality assurance yields
three key benefits.
1. Improved predictive accuracy — the method enhances defect prediction across
the NASA MDP and PROMISE datasets.
2. Enhanced robustness — performance on the NAB benchmark confirms stability
under domain drift.
3. Trustworthy predictions — calibration metrics demonstrate that the system pro-
duces reliable probability estimates, which is essential for safety-critical energy applica-
tions.
The proposed QA framework is also conceptually aligned with ISO/IEC 25010 soft-
ware quality evaluation principles [15].
Taken together, these results demonstrate that the proposed framework not only im-
proves predictive accuracy but also establishes a foundation for trustworthy quality as-
surance in real-world IEMS deployments. The proposed framework provides a founda-
tion for further research on self-evolving quality assurance methods for AI-driven intelli-
gent energy management software [18–20].
6. Conclusions
This study has proposed a unified framework for cross-domain adaptation and
trusted quality assurance (QA) in Intelligent Energy Management Software (IEMS). By
combining transfer learning, adversarial domain alignment, and federated aggregation
with QA mechanisms, namely calibration, robustness assessment, and explainability, the
framework addresses the dual challenges of cross-domain generalization and trustworthy
deployment.
Wang Zhihai, A. A. Verlan
90
Experimental validation on the NASA MDP and PROMISE defect datasets, to-
gether with the NAB and UCI Energy benchmarks, confirmed the effectiveness of the
approach. The proposed method achieved an absolute F1-score improvement of 5–10 per-
centage points on defect prediction, an 8-point increase on NAB anomaly detection (from
0,70 to 0,78), and improved calibration with ECE = 0,032 and NLL = 0,18.
By coupling adaptation strategies with formal QA checks, the framework enhances
both predictive accuracy and reliability, providing a principled and practical solution for
the deployment of IEMS in real-world energy management scenarios.
1. Pan, S. J., & Yang, Q. A Survey on Transfer Learning. IEEE Transactions on Knowledge and
Data Engineering, 2010, Vol. 22, no. 10, pp. 1345-1359. DOI: 10.1109/TKDE.2009.191
2. Ganin, Y., & Lempitsky, V. Unsupervised Domain Adaptation by Backpropagation. Proceedings
of the 32nd International Conference on Machine Learning, ICML, 2015, Vol. 37, pp. 1180–1189. Available
at: https://proceedings.mlr.press/v37/ganin15.html. (accessed 10.08.2025).
3. Hall, T., Beecham, S., Bowes, D., Gray, D., & Counsell, S. A Systematic Literature Review on
Fault Prediction Performance in Software Engineering. IEEE Transactions on Software Engineering, 2012,
Vol. 38, No. 6, pp. 1276–1304. DOI: 10.1109/TSE.2011.103.
4. McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A. Communication-efficient
learning of deep networks from decentralized data. in Proc. 20th Int. Conf. Artificial Intelligence and Sta-
tistics (AISTATS), PMLR, 2017, Vol. 54, pp. 1273-1282. Available at: https://proceed-
ings.mlr.press/v54/mcmahan17a.html. (accessed 12.08.2025).
5. Kairouz, P., & McMahan, H. B. Advances and Open Problems in Federated Learning. Founda-
tions and Trends in Machine Learning, 2021, Vol. 14, Iss. 1–2, pp. 1–210. DOI: 10.1561/2200000083.
6. Guo C., Pleiss G., Sun Y. and Weinberger K.Q., On calibration of modern neural networks, Pro-
ceedings of the 34th International Conference on Machine Learning (ICML), 2017, pp. 1321–1330. DOI:
10.48550/arXiv.1706.04599.
7. Hendrycks, D., & Dietterich, T. Benchmarking Neural Network Robustness to Common Corrup-
tions and Perturbations. International Conference on Learning Representations (ICLR), New Orleans, Lou-
isiana, United States, 2019. DOI: 10.48550/arXiv.1903.12261.
8. Lundberg S.M. and Lee S.-I., A unified approach to interpreting model predictions, Advances in
Neural Information Processing Systems (NeurIPS), 2017, Vol. 30, pp. 4765–4774. DOI:
10.48550/arXiv.1705.07874.
9. Ribeiro M.T., Singh S. and Guestrin C., Why should I trust you?: Explaining the predictions of
any classifier, Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining (KDD), 1135–1144, 2016. DOI: 10.1145/2939672.2939778.
10. Felderer, M., & Ramler, R. Quality Assurance for AI-Based Systems: Overview and Challenges
(Introduction to Interactive Session). Software Quality: Future Perspectives on Software Engineering Qual-
ity, 13th International Conference SWQD, Vienna, Austria, 2021, pp. 33-42. DOI: 10.1007/978-3-030-
65854-0_3
11. Lavin, A., & Ahmad, S. Evaluating Real-Time Anomaly Detection Algorithms – The Numenta
Anomaly Benchmark. Proceedings of the 14th IEEE International Conference on Machine Learning and
Applications (ICMLA), 2015, pp. 38–44. DOI: 10.1109/ICMLA.2015.141.
12. Jureczko, M., & Madeyski, L. Towards Identifying Software Project Clusters with Respect to
Defect Prediction. Proceedings of the 6th International Conference on Predictive Models in Software En-
gineering, PROMISE, Timishoara, Romania, 2010, article No. 9. pp. 1–10. DOI:
10.1145/1868328.1868342.
13. S. Kim, E. J. Whitehead, and Y. Zhang, "Classifying Software Changes: Clean or Buggy?," in
IEEE Transactions on Software Engineering, Vol. 34, No. 2, pp. 181-196, March-April 2008, DOI:
10.1109/TSE.2007.70773.
14. Dua, D. and Graff, C. UCI Machine Learning Repository. University of California, Irvine, School
of Information and Computer Sciences, 2019. Available at: https://archive.ics.uci.edu/datasets.
Сross-domain adaptation and trusted quality assurance of intelligent energy management software
ISSN 1560-9189 Реєстрація, зберігання і обробка даних. 2026. Т. 28, № 2 91
15. International Organization for Standardization. (2023). ISO/IEC 25010:2023 – Systems and soft-
ware engineering — Systems and software Quality Requirements and Evaluation (SQuaRE) — Product
quality model. ISO. Available at: https://www.iso.org/standard/78176.html#lifecycle.
16. Breiman L., Random forests, Machine Learning, 2001, Vol. 45, No. 1, P. 5–32. DOI:
10.1023/A:1010933404324.
17. Hochreiter S. and Schmidhuber J., Long short-term memory, Neural Computation, 1997, Vol. 9,
No. 8, pp. 1735–1780. DOI: 10.1162/neco.1997.9.8.1735.
18. Verlan, A., & Zhihai, W. Toward self-evolving quality assurance frameworks for AI-driven in-
telligent energy management software. Radioelectronic and Computer Systems, 2026, No. 1, pp. 163–178.
DOI: 10.32620/reks.2026.1.11.
19. Verlan, A. А., Zhihai, W., & Yunhai, Z. Enhancing Reliability of Energy Management Software
Through Predictive Modeling and Automated Repair. Connectivity, 2025, Vol. 6, pp. 95–102, DOI:
10.31673/2412-9070.2025.061212.
20. Verlan, A. A., Zhi Hai, W., & Chen, C. Modelling the quality assurance of AI based intelligent
energy management software. Modern Information Security, 2025, Vol. 3 (63), pp. 199–204. DOI:
10.31673/2409-7292.2025.030192.
Received 12.04.2026
Accepted 19.05.2026
Published 17.06.2026
|
| id | drspiprikievua-article-363165 |
| institution | Data Recording, Storage & Processing |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2026-06-18T01:00:39Z |
| publishDate | 2026 |
| publisher | Інститут проблем реєстрації інформації НАН України |
| record_format | ojs |
| resource_txt_mv | drspiprikievua/42/df9646ddeaa251420b96b3f42c4b4642.pdf |
| spelling | drspiprikievua-article-3631652026-06-17T17:50:04Z Сross-domain adaptation and trusted quality assurance of intelligent energy management software Крос-доменна адаптація та забезпечення довіреної якості інтелек-туального програмного забезпечення для управління енергетикою Жихай, Ванг Верлань, А. А. Intelligent Energy Management Software (IEMS); Cross-Domain Adaptation; Transfer Learning; Domain-Adversarial Training; Trustworthy Artificial Intelligence; Software Quality Assurance; Calibration; Explainability інтелектуальне програмне забезпечення для управління енергетикою (IEMS), крос-доменна адаптація, навчання з перенесенням, змагальне вирівнювання доменів, до-вірчий штучний інтелект, забезпечення якості програмного забезпечення, калібрування, пояснюваність, виявлення аномалій, енергетичні мережі Intelligent Energy Management Software must ensure reliable operation across heterogeneous domains where data distributions and system environments frequently vary. This paper introduces a cross-domain adaptation and trusted quality assurance framework that combines transfer learning, adversarial domain alignment, and calibration based reliability assessment, validated on NASA MDP, PROMISE, NAB, and UCI energy benchmarks. Compared with strong baselines such as Random Forest, Convolutional Neural Networks, and Gated Recurrent Units, the proposed method achieves consistent improvements, yielding absolute F1 score gains of 5 to 10 points on defect prediction (NASA MDP, PROMISE) and an increase of 8 points on anomaly detection (NAB, from 0.70 to 0.78), while reducing Expected Calibration Error to 0.032 (a 22 to 42 percent reduction relative to Bayesian CNN baselines) and Negative Log Likelihood to 0.18, thereby demonstrating that integrating cross-domain adaptation with rigorous quality assurance mechanisms significantly enhances both predictive performance and reliability in real world IEMS deployments. Інтелектуальне програмне забезпечення для управління енергетикою (IEMS) відіграє дедалі важливішу роль у сучасних смарт-мережах, розподілених енергетичних системах і платформах оптимізації енергоспоживання. Однією із ключових проблем під час практичного розгортання таких систем є суттєва варіативність розподілів даних, умов експлуатації, конфігурацій сенсорів і прог-рамної інфраструктури в різних середовищах (доменах). Традиційні підходи до забезпечення якості (QA),що орієнтовані на статичні та гомогенні умови, часто призводять до деградації продуктивності моделей штучного інтелекту при їхньому перенесенні між доменами. Це створює суттєві ризики для надійності, стабільності роботи енергомереж і безпеки прийняття рішень. У роботі представлено уніфіковану рамкову структуру, що поєднує механізми крос-доменної адаптації і забезпечення довіреної якості для IEMS. Запропонований підхід інтегрує методи навчання з перенесенням (transfer learning), змагального вирівнювання доменів (adversarial domain alignment) і федеративного агрегування для подолання розбіжностей у розподілах даних. Крім того, структура включає рівень довірчого контролю якості, який охоплює оцінку калібрування (Expected Calibration Error, Negative Log Likelihood), аналіз робастності та пояснюваності, що відповідає принципам довірчого штучного інтелекту (Trustworthy AI) та стандартам якості програмного забезпечення (ISO/IEC 25010). Експериментальну валідацію проведено на загальнодоступних наборах даних, включаючи NASA MDP, PROMISE (прогнозування дефектів), NAB та UCI Energy (виявлення аномалій і прогнозування навантаження). Порівняно із сильними базовими моделями (випадковий ліс, згорткові нейронні мережі, керовані рекурентні блоки) запропонований метод демонструє стабільне покращення: абсолютний приріст F1-показника на 5–10 відсоткових пунктів для прогнозування дефектів і збільшення метрики виявлення аномалій з 0,70 до 0,78 на NAB. Важливо, що досягнуто очікуваної помилки калібрування (ECE) на рівні 0,032 (що відповідає зниженню на 22–42 % порівняно з баєсовими згортковими мережами) та зменшення від’ємної логарифмічної правдоподібності (NLL) до 0,18. Це підтверджує, що модель забезпечує надійні ймовірнісні оцінки, критично важливі для прийняття рішень у безпекокритичних енергетичних застосунках. Таким чином, запропонована рамкова структура не лише підвищує точність прогнозів в умовах зміни доменів, але й закладає основу для практичного розгортання сертифікованих, пояснюваних і стійких систем IEMS у реальних сценаріях управління енергетикою. Результати роботи відкривають шлях до створення самоеволюційних методів забезпечення якості для наступного покоління енергетичного програмного забезпечення. Табл.: 5. Іл.: 5. Бібліогр.: 20 найм. Інститут проблем реєстрації інформації НАН України 2026-06-17 Article Article application/pdf https://drsp.ipri.kiev.ua/article/view/363165 10.35681/1560-9189.2026.28.2.363165 Data Recording, Storage & Processing; Vol. 28 No. 2 (2026); 79-91 Регистрация, хранение и обработка данных; Том 28 № 2 (2026); 79-91 Реєстрація, зберігання і обробка даних; Том 28 № 2 (2026); 79-91 1560-9189 en https://drsp.ipri.kiev.ua/article/view/363165/350529 Авторське право (c) 2026 Реєстрація, зберігання і обробка даних |
| spellingShingle | Intelligent Energy Management Software (IEMS) Cross-Domain Adaptation Transfer Learning Domain-Adversarial Training Trustworthy Artificial Intelligence Software Quality Assurance Calibration Explainability Жихай, Ванг Верлань, А. А. Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title | Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title_alt | Крос-доменна адаптація та забезпечення довіреної якості інтелек-туального програмного забезпечення для управління енергетикою |
| title_full | Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title_fullStr | Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title_full_unstemmed | Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title_short | Сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| title_sort | сross-domain adaptation and trusted quality assurance of intelligent energy management software |
| topic | Intelligent Energy Management Software (IEMS) Cross-Domain Adaptation Transfer Learning Domain-Adversarial Training Trustworthy Artificial Intelligence Software Quality Assurance Calibration Explainability |
| topic_facet | Intelligent Energy Management Software (IEMS) Cross-Domain Adaptation Transfer Learning Domain-Adversarial Training Trustworthy Artificial Intelligence Software Quality Assurance Calibration Explainability інтелектуальне програмне забезпечення для управління енергетикою (IEMS) крос-доменна адаптація навчання з перенесенням змагальне вирівнювання доменів до-вірчий штучний інтелект забезпечення якості програмного забезпечення калібрування пояснюваність виявлення аномалій енергетичні мережі |
| url | https://drsp.ipri.kiev.ua/article/view/363165 |
| work_keys_str_mv | AT žihajvang srossdomainadaptationandtrustedqualityassuranceofintelligentenergymanagementsoftware AT verlanʹaa srossdomainadaptationandtrustedqualityassuranceofintelligentenergymanagementsoftware AT žihajvang krosdomennaadaptacíâtazabezpečennâdovírenoíâkostííntelektualʹnogoprogramnogozabezpečennâdlâupravlínnâenergetikoû AT verlanʹaa krosdomennaadaptacíâtazabezpečennâdovírenoíâkostííntelektualʹnogoprogramnogozabezpečennâdlâupravlínnâenergetikoû |