AI Network Coverage Planning System Development

Development of an AI system for network coverage planning

Radio coverage planning is an iterative task: deploying the minimum number of base stations to provide a given coverage within a limited budget. AI replaces months of manual analysis by RF engineers with weeks of automated planning.

Radio coverage forecasting

Radio signal propagation models:

Physical models (Okumura-Hata, COST 231) provide a basic signal loss forecast. Limitation: they do not take into account local topography, buildings, or vegetation.

ML correction of the physical model:

import numpy as np
import pandas as pd
from lightgbm import LGBMRegressor

class PropagationMLModel:
    """
    ML-поправки к физической модели распространения сигнала.
    Данные: измерения уровня сигнала от drive test + CQT
    """

    def build_features(self, measurement_df, dem_raster, building_footprints):
        """
        measurement_df: GPS + RSRP/RSSI измерения
        dem_raster: цифровая модель рельефа (SRTM/Copernicus 30m)
        building_footprints: OSM или кадастр
        """
        df = measurement_df.copy()

        # Рельефные признаки
        df['elevation'] = self._sample_dem(df[['lat', 'lon']], dem_raster)
        df['elevation_diff'] = df['elevation'] - self._sample_dem_bs(df['bs_id'])
        df['terrain_roughness'] = self._calculate_roughness(df[['lat', 'lon']], dem_raster, radius_m=500)

        # Застройка
        df['building_density_500m'] = self._building_coverage(df[['lat', 'lon']],
                                                               building_footprints, radius=500)
        df['mean_building_height_200m'] = self._mean_height(df[['lat', 'lon']],
                                                             building_footprints, radius=200)

        # Геометрия от БС
        df['distance_to_bs'] = self._haversine_distance(df[['lat', 'lon']], df['bs_location'])
        df['angle_to_bs'] = self._bearing(df[['lat', 'lon']], df['bs_location'])
        df['los_probability'] = self._estimate_los(df, building_footprints)

        # Физическая модель как базовая фича
        df['okumura_hata_pred'] = self._okumura_hata(df)

        return df

    def train(self, features_df, target='rsrp_dbm'):
        feature_cols = [c for c in features_df.columns
                       if c not in [target, 'lat', 'lon', 'timestamp', 'bs_id']]
        self.model = LGBMRegressor(n_estimators=500, learning_rate=0.03, num_leaves=64)
        self.model.fit(features_df[feature_cols], features_df[target])
        return self

Drive Test replacement with AI predictions:

The traditional approach: drive along all the streets with measuring equipment. The AI approach: an ML model, based on trained data, predicts the RSRP at any point → reducing the volume of drive tests by 60–75%.

Optimization of base station placement

Integer Programming for Site Selection:

import pulp

def optimize_site_selection(
    candidate_sites,   # возможные площадки с характеристиками
    coverage_predictions,  # матрица: site × pixel → predicted RSRP
    demand_map,        # карта спроса на трафик
    budget_usd,
    rsrp_threshold=-95  # минимально допустимый RSRP в dBm
):
    prob = pulp.LpProblem("site_selection", pulp.LpMaximize)

    # Бинарные переменные: строить ли сайт i
    build = [pulp.LpVariable(f"build_{i}", cat='Binary') for i in range(len(candidate_sites))]

    # Переменные покрытия: покрыт ли пиксель j
    covered = [pulp.LpVariable(f"covered_{j}", cat='Binary')
               for j in range(coverage_predictions.shape[1])]

    # Objective: максимизировать взвешенное покрытие (по трафику)
    prob += pulp.lpSum(demand_map[j] * covered[j] for j in range(len(covered)))

    # Бюджет
    prob += pulp.lpSum(candidate_sites[i]['cost'] * build[i]
                       for i in range(len(candidate_sites))) <= budget_usd

    # Пиксель покрыт если хотя бы один сайт обеспечивает нужный RSRP
    for j in range(len(covered)):
        covering_sites = [i for i in range(len(candidate_sites))
                         if coverage_predictions[i][j] >= rsrp_threshold]
        if covering_sites:
            prob += covered[j] <= pulp.lpSum(build[i] for i in covering_sites)
        else:
            prob += covered[j] == 0  # этот пиксель нельзя покрыть

    prob.solve(pulp.PULP_CBC_CMD(msg=0, timeLimit=120))

    selected_sites = [i for i, b in enumerate(build) if b.value() > 0.5]
    return selected_sites

5G planning

Millimeter waves (mmWave, 26/28 GHz):

mmWave - high throughput, but very limited coverage: - Radio Line-of-Sight requirement: any obstacle (tree, person) = significant losses - Inter-site distance: 150-300 m vs. 500-1000 m for Sub-6GHz

ML tasks specific to 5G mmWave: - Blockage prediction: the probability of LoS loss on a specific route - Beam management: selecting the optimal beam from 64 possible ones based on the channel angular profile - Dual connectivity: when to switch from 5G to 4G LTE (fallback)

Small cells & HetNet:

Outdoor small cells (small cells, femtocells) + macro networks (HetNet): - ML clustering of high-demand points → optimal positions for small cells - Interference coordination: AI manages power/frequency to minimize inter-cell interference

Coverage analytics

Coverage gap analysis:

• Subscriber complaints + GPS → map of problem areas - Neural network links complaints to network parameters → precise causes - Automatic prioritization: where improvements will yield the greatest increase in NPS

Development time: 3-5 months for an AI coverage planning system with ML RSRP prediction and base station placement optimization.