PREDICTIVE MODELING AND OPTIMIZATION OF RADIO COVERAGE IN NEXT-GENERATION MOBILE NETWORKS

УДК 621.396.96.

ABSTRACT

Accurate radio coverage prediction and base station placement optimization remain coupled unsolved problems in fifth-generation cellular network planning. Empirical path loss models produce errors exceeding 15 dB in morphologically complex environments, and published machine learning approaches improve prediction fidelity yet fix model parameters after offline training without any mechanism for revision against operational evidence. This article presents an adaptive framework in which a convolutional neural network propagation model is continuously updated against live network measurements through a closed-loop feedback architecture, coupled to a demand-weighted site candidate generator, a multi-criteria feasibility assessor, and a dynamic reallocation controller. Evaluation across dense urban, rural, and suburban densification scenarios yielded prediction error reductions of 3 to 6 dB relative to non-adapted baselines, planning cycle durations of 14 to 22 days against 55 to 70 days for manual workflows, and a base station count reduction of 28 to 35 percent at equivalent coverage thresholds.

АННОТАЦИЯ

Точное прогнозирование радиопокрытия и оптимизация размещения базовых станций остаются взаимосвязанными нерешёнными задачами при планировании сотовых сетей пятого поколения. Эмпирические модели потерь распространения радиосигнала дают погрешности, превышающие 15 дБ, в морфологически сложных средах, а опубликованные подходы на основе машинного обучения повышают точность прогнозирования, однако фиксируют параметры модели после офлайн-обучения, не предусматривая механизмов их корректировки на основе данных эксплуатации. В настоящей статье представлена адаптивная структура, в которой модель распространения радиосигнала на основе свёрточной нейронной сети непрерывно обновляется по результатам измерений в реальной сети через архитектуру замкнутой обратной связи, дополненную генератором кандидатных площадок с учётом весового распределения спроса, многокритериальным модулем оценки реализуемости и контроллером динамического перераспределения. Оценка на сценариях плотной городской застройки, сельской местности и пригородного уплотнения показала снижение ошибки прогнозирования на 3–6 дБ по сравнению с неадаптивными базовыми моделями, длительность цикла планирования 14–22 дня против 55–70 дней при ручных рабочих процессах, а также сокращение числа базовых станций на 28–35 % при сохранении эквивалентных порогов покрытия.

Keywords:  radio coverage prediction, propagation modeling, machine learning, 5G network planning, base station placement, adaptive modeling, geospatial data fusion.

Ключевые слова: прогнозирование радиопокрытия, моделирование распространения радиоволн, машинное обучение, планирование сетей 5G, размещение базовых станций, адаптивное моделирование, объединение геопространственных данных.

Introduction

Network planning for cellular infrastructure requires reconciling two demands with partially conflicting resource requirements: physical accuracy in propagation modeling and computational tractability at metropolitan deployment scale. Deterministic simulation methods achieve engineering-grade precision in representing diffraction, reflection, and scattering, yet their computational cost has precluded iterative use across the hundreds of candidate sites a full urban deployment entails. Fast empirical formulas scale to large regions at low computational cost but encode morphological assumptions derived from specific measurement campaigns; when applied to environments with materially different building height distributions or terrain profiles, these assumptions produce systematic prediction errors that the model's functional form cannot correct internally [14]. For earlier mobile generations operating at sub-2 GHz frequencies with generous macro-cell link budgets, prediction errors on the order of 8 to 12 dB were operationally tolerable, as link budget margins absorbed residual model inaccuracy without measurable service degradation [10]. Fifth-generation new radio reduces this tolerance substantially. Massive MIMO arrays, spectrum above 3.5 GHz, and ultra-dense small cell overlays reduce the permissible prediction error margin to a level at which a 3 dB overestimate of received power at a cell-edge grid point produces a coverage outage zone in the deployed network [13].

Base station placement introduces a further optimization problem of considerable combinatorial complexity. Identifying locations across a large metropolitan area requires evaluating candidate sites against geospatial, regulatory, structural, logistical, and financial constraints simultaneously, and classical workflows accomplish this through expert judgment supported by drive-test campaigns, producing a deployment plan that is static from the moment of approval and can require several months to finalize [9]. Eisenblätter et al. (2004) established that propagation model quality and optimizer performance are tightly coupled, with improvements to either component in isolation producing diminishing returns. In practice, however, the two components have continued to be developed and applied as independent sequential procedures [4]. The absence of a feedback pathway from operational network performance to the planning model is a structural feature of all current planning workflows. Erceg et al. (2001) reported statistically significant deviations in path loss exponent and shadowing standard deviation between reference and target environments sharing the same morphological classification, indicating that propagation model parameters are site-specific to a degree that a single offline calibration procedure cannot fully account for [5]. A further complication arises from the spatial distribution of available measurement data: observational density is highest in sub-regions with established network coverage, whereas coverage-deficient zones, for which model correction would yield the greatest planning benefit, contribute disproportionately few measurement samples to any calibration dataset.

The framework described in this article integrates propagation modeling and deployment optimization into a continuously adapting system. A convolutional neural network trained in sequential stages and updated against live measurements generates probabilistic coverage estimates over a multi-layer geospatial region representation. A demand-weighted greedy optimizer identifies candidate locations subject to configurable constraints on site count and spacing. A feasibility assessor scores candidates against four categories of practical constraint and produces an actionable ranked plan. A dynamic reallocation controller attributes post-deployment performance discrepancies to their causal class and triggers targeted corrections to the model or the deployment plan as appropriate.

Materials and Methods

The Okumura-Hata model expresses median path loss as a function of carrier frequency, antenna heights, and propagation distance, with additive correction terms for urban, suburban, and open morphological classes, derived from Tokyo measurement campaigns at 150 MHz to 1.5 GHz [6]. The COST 231 extension raised the frequency ceiling to 2 GHz and introduced a dense-urban coefficient. In both models, regression coefficients encode the propagation statistics of the original measurement environments, and independent validation datasets for deep urban canyons report prediction errors exceeding 15 dB for Okumura-Hata and above 10 dB for COST 231 [8]. At 5G frequencies above 3.5 GHz, where rooftop diffraction and facade reflection constitute primary, the scalar distance-frequency parameterization of these models does not provide sufficient degrees of freedom to represent geometry-dependent attenuation variability [14]. Ray-tracing addresses this limitation by tracing electromagnetic rays through three-dimensional building databases and yields 3 to 6 dB root mean square errors in environments with accurate building data [16]. Several hours of server computation are required per candidate configuration over a 10 km² urban scene, which renders iterative optimization at metropolitan scale computationally infeasible with current hardware [3].

Neural network models trained on local measurement data achieve prediction accuracy approaching that of ray-tracing at substantially lower computational cost. Feed-forward architectures reduce prediction error by 2 to 5 dB relative to Okumura-Hata on matched environments, and convolutional architectures trained on geospatial raster inputs learn spatially varying attenuation corrections that account for canyon geometry and forest clutter without explicit physical modeling of those mechanisms [15]. Transfer learning between morphologically similar reference and target environments reduces the volume of target-region measurement data required for adequate model specialization [7]. Across all published models in this category, parameters are determined during offline training and receive no subsequent updates from operational network measurements; prediction error therefore accumulates as the propagation environment evolves through construction activity, seasonal vegetation change, and post-deployment antenna reconfiguration.

The base station placement problem is a weighted maximum coverage problem, NP-hard in the general case [2]. Greedy incremental selection yields an approximation ratio of (1−1/e) ≈ 0.63 relative to the global optimum when the coverage objective is monotone submodular, a property that holds under standard signal threshold models by virtue of the diminishing returns structure of overlapping coverage areas. Metaheuristic approaches including simulated annealing and genetic algorithms may exceed this bound on specific problem instances but provide no general polynomial-time approximation guarantee. Eisenblätter et al. (2004) found that a 3 dB improvement in propagation model accuracy reduced required site count more than any algorithmic modification tested on the same dataset [4]. Academic placement formulations have consistently omitted the practical feasibility constraints governing real deployments; Mishra (2007) reported that 20 to 30 percent of coverage-optimal candidate locations fail regulatory, structural, logistical, or cost criteria in typical urban settings [9].

The proposed framework comprises five modules operating continuously across planning, deployment, and operational phases. The geospatial data ingestion module processes digital elevation models at 5 to 30 m resolution, building footprint and height data from municipal GIS, drive-test and crowdsourced signal measurements, regulatory constraint layers, and privacy-aggregated user activity records into time-varying demand surfaces; a fusion pipeline aligns all inputs to a common grid, fills missing values by interpolation, and normalizes features into a multi-channel region representation. The adaptive coverage modeling engine applies a convolutional neural network with three to six layers and 32 to 256 filters per layer to extract spatially structured feature maps from the region representation; a prediction subnetwork combines these features with base station configuration parameters (antenna height, azimuth, downtilt, transmit power, carrier frequency, and bandwidth) and outputs at each grid point the parameters of a Gaussian distribution over received signal strength, parameterized by mean μ_i and variance σ_i, with the probabilistic representation allowing prediction confidence to enter downstream candidate ranking. Training proceeds in three stages: initial training on reference-region historical data; regional fine-tuning on target-region data under reduced learning rates and constrained parameter updates; and online adaptation against streaming operational measurements, regularized by an L2 penalty on parameter changes relative to the fine-tuned state.

The three-stage training sequence is motivated by the empirical finding that direct online adaptation from random initialization on sparse target-region data leads to catastrophic forgetting of propagation relationships learned from the larger reference-region dataset. Initial training on reference-region data establishes a parameter configuration encoding general geospatial-to-path-loss relationships across diverse morphologies; fine-tuning then specializes those parameters to the target region under a constrained update regime that preserves generalizable features while correcting for local deviations; and the L2 regularization applied during online adaptation prevents overfitting to the uneven spatial distribution of live measurement data at operational timescale.

The site candidate generator computes a demand-weighted coverage deficit at each grid point:

where   is predicted coverage probability,    is the minimum acceptable threshold, and   is the normalized demand weight with  . High-deficit points are grouped into priority zones by a region-growing algorithm, and the greedy selection module iteratively identifies locations maximizing the reduction in   subject to site count and spacing constraints, with a local perturbation stage to reduce grid quantization sensitivity. The feasibility assessor evaluates each candidate against regulatory, structural, logistical, and cost criteria, and ranks candidates by the product of composite feasibility score and expected coverage contribution. The dynamic reallocation controller monitors operational performance indicators from deployed sites, identifies sub-regions where observed performance diverges from predictions beyond a prescribed threshold, and determines the causal class of each discrepancy: propagation model errors trigger regularized parameter updates to the coverage modeling engine, and configuration-induced anomalies trigger antenna parameter adjustment recommendations without modifying the propagation model.

The five modules operate in a defined sequential cycle during the planning phase and in a continuous monitoring cycle during network operation. At planning time, the ingestion module produces the region representation, which is passed to the coverage modeling engine to generate a full-region coverage probability map; the candidate generator processes this map against the current demand surface to produce an ordered candidate set; and the feasibility assessor filters and ranks that set into a deployment plan. During network operation, the reallocation controller receives performance reports from deployed sites at configurable intervals, compares observed indicators against the coverage modeling engine's predictions for the same grid sub-regions, and routes the resulting discrepancy records either to the model adaptation submodule or to the configuration recommendation interface depending on the causal class assigned by the attribution module.

Three scenarios were evaluated against four baselines under identical coverage thresholds. The dense urban scenario covered 150 to 250 km² with buildings from 20 to over 150 m, using drive-test data along approximately 800 km of road at 3.5 GHz. The rural scenario addressed 1,800 to 2,200 km² of agricultural and forested terrain with towns of 500 to 5,000 residents, with crowdsourced measurements as the primary data source. The suburban densification scenario covered a region of 400 to 600 km² with an existing 47-site 5G NR macro network, using operational performance records for model calibration. Baselines were a prior-art machine learning planning system, a COST 231-based tool, an Okumura-Hata-based tool, and a manual engineering workflow.

Results

In the dense urban scenario, target-region fine-tuning reduced mean prediction error in the central business district by 3 to 6 dB relative to the non-adapted model. The improvement was spatially non-uniform: the largest reductions occurred in deep canyon zones and near irregular rooftop geometries not represented in the reference-region training data, where the non-adapted model systematically underestimated facade reflection attenuation, consistent with the bias correction mechanism described by Thrane et al. (2020) [15].

In the rural scenario, elevated regularization during fine-tuning produced wider predictive variance over forested sub-regions, accurately reflecting sparse measurement support. Candidate sites were concentrated near elevated terrain features adjacent to road corridors, where coverage radii at 700 MHz reached 8 to 15 km against 3 to 5 km from topographically lower locations with equivalent antenna heights. This spatial distribution of candidates is a direct consequence of the demand weighting in the objective function, as road-corridor grid points carry substantially higher w_i values than adjacent agricultural land [5].

In the suburban densification scenario, calibration against 47 macro cell operational records enabled the model to identify sector-level load imbalances in commercial corridors that a coverage-only formulation would not detect. The combined coverage-capacity deficit formulation identified 12 to 15 percent more high-priority candidate locations than a coverage-only model applied to the same sub-region [1]. Table 1 summarizes principal planning efficiency metrics for the urban scenario.

Table 1. 

Comparative deployment metrics, urban evaluation scenario

The approximately 15 percent reduction in site count relative to the prior-art machine learning system is attributable to online model adaptation: in the absence of a feedback mechanism, that system accumulated prediction errors as environmental conditions diverged from the training snapshot, producing suboptimal candidate selection in affected sub-regions. The 28 to 35 percent reduction relative to manual practice reflects the combined contribution of improved prediction accuracy, demand-weighted candidate selection, and integrated feasibility assessment. In the urban scenario, 18 percent of coverage-optimal candidates required additional structural review and 7 percent were excluded on regulatory grounds; composite feasibility scoring eliminated the associated manual review latency from the planning cycle.

Post-deployment monitoring in the urban scenario identified three underground rail sub-regions where observed received power fell 8 to 12 dB below predictions. Attribution analysis identified reinforced concrete tunnel wall attenuation as the source, a mechanism absent from both training datasets. The controller generated targeted model update data from measurements in the affected zones, applied regularized gradient updates through the adaptation submodule, and identified indoor and tunnel-entry small cell candidate locations within a single operational cycle. In the suburban scenario, two small cells exhibited inter-sector interference levels 4 dB above predictions. The attribution module classified this as a configuration-induced anomaly, on the basis that the discrepancy correlated with macro cell overlap geometry and was not associated with any geospatial feature indicative of a missing propagation mechanism; antenna downtilt and azimuth adjustments were recommended for the implicated macro sectors without triggering propagation model updates.

Discussion

The site count reduction relative to the prior-art machine learning system sharing the same neural network architecture is explained by the online adaptation mechanism alone. Incremental improvements to convolutional architecture design or offline training dataset size produce accuracy gains that erode at the rate of environmental change once the model operates without a post-deployment correction mechanism. Krijestorac et al. (2021) found that transfer learning outperformed deeper static architectures on unseen target environments, consistent with the interpretation that representational capacity is not the principal constraint on planning accuracy in current systems [7]. The primary source of prediction error accumulation is the assumption of environmental stationarity implicit in single-pass offline calibration, and the closed-loop design addresses this at the architectural level.

Extension to millimeter-wave frequencies above 24 GHz involves propagation mechanisms, specifically molecular absorption, near-field blockage, and rapid spatial decorrelation, that the terrain and land-use features used in the present evaluation do not represent at adequate resolution [13]. Prediction accuracy at those frequencies requires building facade material properties and sub-meter street geometry at a resolution that most municipal GIS databases do not currently provide, and the three-stage training scheme requires validation with millimeter-wave measurement data before its efficiency can be generalized to that frequency regime. The use of privacy-aggregated user activity data for demand surface construction is subject to regulatory governance requirements that vary across jurisdictions and depend on the spatial resolution at which demand data is processed; these requirements are context-specific and require separate assessment for each deployment environment.

Conclusion

The results reported across three deployment scenarios of contrasting morphological character provide empirical support for the hypothesis that the primary source of prediction error accumulation in cellular network planning is not inadequate model expressiveness but the structural absence of a mechanism for post-deployment model correction. The framework reported in this article addresses this by coupling the propagation model, the optimizer, the feasibility assessor, and the post-deployment monitoring system into a single adaptive loop. Online adaptation accounts for half of the 28 to 35 percent site count reduction relative to manual practice observed across evaluation scenarios, with demand-weighted candidate selection and integrated feasibility scoring contributing the remainder. Across all three evaluation scenarios, the closed-loop design maintained prediction alignment with the operational environment without requiring new measurement campaigns or manual re-initialization of the planning process.

References:

1. Andrews J. G. What will 5G be? / J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, J. C. Zhang // IEEE Journal on Selected Areas in Communications. – 2014. – № 6 (32). – P. 1065–1082.
2. Amaldi E. Planning UMTS base station location: Optimization models with power control and algorithms / E. Amaldi, A. Capone, F. Malucelli // IEEE Transactions on Wireless Communications. – 2003. – № 5 (2). – P. 939–952.
3. Degli-Esposti V. An advanced field prediction model including diffuse scattering / V. Degli-Esposti, D. Guiducci, A. de'Marsi, P. Azzi, F. Fuschini // IEEE Transactions on Antennas and Propagation. – 2004. – № 7 (52). – P. 1717–1728.
4. Eisenblätter A. Modeling feasible network configurations for UMTS / A. Eisenblätter, A. Fügenschuh, T. Koch, A. M. C. A. Koster, A. Martin, T. Pfender, O. Wegel, R. Wessäly // Telecommunications Network Design and Management. – New York : Springer, 2004. – P. 1–22.
5. Erceg V. An empirically based path loss model for wireless channels in suburban environments / V. Erceg, L. J. Greenstein, S. Y. Tjandra, S. R. Parkoff, A. Gupta, B. Kulic, A. A. Julius, R. Bianchi // IEEE Journal on Selected Areas in Communications. – 2001. – № 7 (17). – P. 1205–1211.
6. Hata M. Empirical formula for propagation loss in land mobile radio services / M. Hata // IEEE Transactions on Vehicular Technology. – 1980. – № 3 (29). – P. 317–325.
7. Krijestorac H. Spatial signal strength prediction using 3D maps and deep learning / H. Krijestorac, S. Hanna, D. Cabric // Proceedings of the IEEE ICC 2021. – IEEE, 2021. – P. 1–6.
8. Mawjoud S. A. Path loss propagation model prediction for GSM network planning / S. A. Mawjoud // International Journal of Computer Applications. – 2013. – № 7 (84). – P. 30–33.
9. Mishra A. R. Advanced Cellular Network Planning and Optimisation. // Chichester : Wiley. – 2007. – 547 p.
10. Molisch A. F. Wireless Communications. – 2nd ed. // Chichester : Wiley–IEEE Press. – 2011. – 884 p.
11. Nemhauser G. L. An analysis of approximations for maximizing submodular set functions / G. L. Nemhauser, L. A. Wolsey, M. L. Fisher // Mathematical Programming. – 1978. – № 1 (14). – P. 265–294.
12. Raisanen L. Comparison and evaluation of multiple objective genetic algorithms for the antenna placement problem / L. Raisanen, R. M. Whitaker // Mobile Networks and Applications. – 2005. – № 1–2 (10). – P. 79–88.
13. Rappaport T. S. Millimeter wave mobile communications for 5G cellular: It will work! / T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, F. Gutierrez // IEEE Access. – 2015. – № 1. – P. 335–349.
14. Saunders S. R. Antennas and Propagation for Wireless Communication Systems. – 2nd ed. // Chichester : Wiley. – 2007. – 546 p.
15. Thrane J. Model-aided deep learning method for path loss prediction in mobile communication systems at 2.6 GHz / J. Thrane, D. Zibar, H. L. Christiansen // IEEE Access. – 2020. – № 8. – P. 7925–7936.
16. Yun Z. Ray tracing for radio propagation modeling: Principles and applications / Z. Yun, M. F. Iskander // IEEE Access. – 2015. – № 3. – P. 1089–1100.