⚡ electrical engineering

Collective Grid: Privacy-Preserved Multi-Operator Energy Sharing Optimization via Federated Energy Prediction

This paper proposes a privacy-preserving framework that enables mobile network operators to collaboratively optimize energy sharing through federated learning-based demand forecasting and mixed-integer linear programming, significantly reducing operational costs and improving efficiency in dense 5G networks.

Original authors: Meysam Masoudi, Tahar Zanouda, Milad Ganjalizadeh, Cicek Cavdar

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Meysam Masoudi, Tahar Zanouda, Milad Ganjalizadeh, Cicek Cavdar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a neighborhood where every house has its own generator, a small solar panel on the roof, and a battery in the garage. Now, imagine that instead of each family trying to figure out the best time to charge their battery or buy cheap electricity on their own, they all decide to work together as a team.

That is essentially what this paper proposes, but instead of houses, we are talking about mobile network towers (the big antennas that give us 5G and internet) owned by different companies (like Verizon, AT&T, or Vodafone).

Here is the story of their new plan, broken down simply:

1. The Problem: Everyone is Playing Solo

Right now, every mobile company manages its own towers independently.

The Waste: Company A might have a tower running on expensive electricity at night, while Company B's tower right next door has a battery full of cheap solar power but doesn't know it.
The Cost: Because they don't talk to each other, they miss out on saving money. It's like two neighbors both buying milk separately at full price, when they could have bought one gallon together and split the cost.
The Privacy Fear: Companies are scared to share their data (like how much traffic their towers handle) because that data is their "secret sauce." They don't want competitors to see their business secrets.

2. The Solution: A "Secret Team" Approach

The authors propose a system called Collective Grid. It's a way for these rival companies to share energy infrastructure (batteries and power lines) without ever revealing their private data to each other.

They use three main "tools" to make this happen:

Tool A: The "Group Brain" (Federated Learning)

This is the coolest part. Imagine a group of students taking a test.

Old Way: Everyone brings their homework to the teacher, the teacher copies it, and then gives it back. (This is bad for privacy; the teacher sees everyone's answers).
New Way (Federated Learning): Each student solves the problem on their own desk. They only send the answer key (the math logic) to the teacher, not their actual homework. The teacher combines all the answer keys to create a "Super Answer Key" that is smarter than any single student's. Then, the teacher sends this Super Key back to everyone.

In this paper, the "students" are the mobile towers. They learn to predict how much electricity they will need based on traffic and weather. They share their "learning patterns" with a central computer, but never share the actual data. This way, the system gets smart at predicting energy needs without anyone spying on the others.

Tool B: The "Smart Shopper" (Optimization)

Once the system knows how much energy will be needed, it acts like a super-smart shopper.

It looks at the price of electricity (which changes every hour).
It looks at how much sun or wind is available.
It looks at how full the shared batteries are.
The Decision: It decides: "Okay, for the next hour, let's use the battery for Tower A because grid power is expensive. But for Tower B, let's plug into the grid because the battery is low."

It does this mathematically to ensure the total bill for everyone is as low as possible.

Tool C: The "Shared Battery" (Infrastructure Sharing)

Instead of every company buying their own giant battery bank for every tower, they build one big shared battery that serves multiple towers from different companies.

Think of it like a community swimming pool. Instead of every family building a $50,000 pool in their backyard, they all chip in to build one giant pool in the neighborhood. Everyone gets to swim, but the cost is shared, and the maintenance is easier.

3. The Results: Why It Works

The researchers tested this idea using real data from a European city with 1,200 towers. Here is what they found:

Money Saved: The more towers that shared the infrastructure, the more money was saved. It's a classic case of "the more the merrier."
The Sweet Spot: The biggest savings happened when companies shared both the batteries and the power lines.
Future Proof: As 5G networks get denser (more towers packed closer together), this sharing model becomes even more valuable.

The Big Picture Analogy

Imagine a group of hikers in a forest.

Without the plan: Each hiker carries their own heavy water bottle, buys expensive water at the trailhead, and gets tired quickly.
With the plan: They form a "caravan." They share a few large water jugs (the shared battery). They use a secret code (Federated Learning) to predict who will get thirsty next without telling the others where they are going. They buy water only when the price is low and fill up the jugs.
The Result: Everyone arrives at the destination with more energy, less fatigue, and a lot more money in their pockets.

In short: This paper shows that if mobile companies stop fighting and start sharing their energy resources using smart, privacy-safe AI, they can save billions of dollars and keep the lights on for our phones more efficiently.

1. Problem Statement

With the expansion of 5G networks, mobile network operators (MNOs) face rising energy consumption due to denser infrastructure and increased data traffic. Currently, MNOs manage energy independently, leading to:

Limited Coordination: Missed opportunities for collective efficiency gains.
High CAPEX: Deploying on-site renewable generation and battery storage for every site individually is capital-intensive.
Data Privacy Barriers: MNOs are reluctant to share raw network data (traffic, configuration, location) due to competitive and regulatory concerns, hindering collaborative optimization.
Unreliable Renewables: Weather-dependent energy sources require dynamic management that is difficult to optimize in isolation.

The core challenge is to enable cross-MNO energy infrastructure sharing (pooling batteries and renewables) to reduce costs while strictly preserving data privacy and accurately predicting heterogeneous energy demands.

2. Methodology

The authors propose a three-module framework designed for co-located MNOs operating in the same geographic area.

A. System Model

Entities: Multiple MNOs ( $M$ ) operating radio sites ( $U$ ). Sites are grouped under shared energy infrastructures ( $I$ ).
Energy Sources: Sites draw power from three sources: On-site Renewables ( $R$ ), the Electrical Grid ( $G$ ), and Shared Battery Storage ( $B$ ).
Battery Dynamics: The model tracks both State of Charge (SoC) and State of Health (SoH). SoH degrades based on charge/discharge cycles and time, affecting the effective battery capacity ( $\Omega$ ) over the long term.
Cost Function: The Total Cost of Ownership (TCO) includes Capital Expenditure (CAPEX) for infrastructure/batteries and Operational Expenditure (OPEX) for electricity purchases and rental fees.

B. Module 1: Federated Learning (FL) for Energy Prediction

To solve the data privacy issue, the framework uses Federated Learning instead of centralizing raw data.

Input Data: Each site uses local time-series data (Key Performance Indicators like throughput and connection success rates) and static Configuration Management (CM) data (hardware models, frequency bands).
Model Architecture: A Multivariate Long Short-Term Memory (LSTM) network is trained locally at each site.
Process:
1. Sites train local models on their specific traffic and weather data.
2. Only model weights (updates) are sent to a central coordinator.
3. The coordinator aggregates updates (using FedAvg) to create a global model.
4. The global model predicts future energy demand ( $\hat{D}_u(t)$ ) for the next control window without ever exposing raw traffic data.

C. Module 2 & 3: Optimization and Orchestration

Once demand is predicted, a Mixed-Integer Linear Program (MILP) is solved to optimize energy sourcing.

Objective: Minimize the total operational cost ( $\Gamma_{Total}$ ) across all sites.
Decision Variables: A binary variable $d_u(t)$ determines the power source for each site at time $t$ (0 = Grid, 1 = Battery).
Constraints:
- SoC Constraints: Battery charge must remain above a minimum threshold ( $SoC_{min}$ ) to handle emergencies.
- SoH Constraints: Battery health must remain above a minimum threshold ( $SoH_{min}$ ) to prevent premature degradation.
- Capacity Constraints: Total discharge cannot exceed available battery capacity.
Inputs to Optimizer: Real-time grid prices, renewable forecasts, predicted demand, and current battery states.
Execution: The system switches energy sources every 15 minutes (control window) based on the optimization result.

3. Key Contributions

Privacy-Preserving Framework: A novel architecture that enables cross-MNO energy sharing using Federated Learning, allowing collaborative demand forecasting without sharing sensitive network data.
Joint Optimization Formulation: A mathematical model that simultaneously optimizes grid purchases, renewable utilization, and shared battery charging/discharging while accounting for battery degradation (SoH) and varying ownership shares ( $\alpha_{iu}$ ).
Comprehensive Evaluation: Validation using real-world operational data from a European region with 1,200 sites across three MNOs, demonstrating the viability of the approach in a 5G context.

4. Simulation Results

The framework was tested using data from 1,200 sites over a 15-year simulation period.

Prediction Accuracy: The FL-based LSTM model achieved a Mean Absolute Error (MAE) of approximately 0.39–0.43 (depending on hyperparameters), proving effective for forecasting energy demand across heterogeneous sites.
Cost Reduction:
- Baseline: Exclusive infrastructure (Grid only) had the highest cumulative cost.
- Sharing Benefits: Sharing infrastructure (without batteries) reduced costs. Adding shared batteries further reduced costs.
- Optimal Scenario: The Shared Infrastructure + Shared Battery configuration yielded the lowest Total Cost of Ownership (TCO) over 15 years.
Impact of Sharing Ratio ( $\zeta$ ): As the percentage of sites participating in sharing increases, annual operational costs decrease significantly.
Long-term Viability: While shared battery setups have higher initial CAPEX, they cross over to become more cost-effective than non-sharing setups within a few years due to reduced OPEX (grid energy savings).
Battery Management: The system successfully maintained SoC above the 10% safety threshold, preventing outages while maximizing the use of stored energy.

5. Significance and Conclusion

This paper addresses a critical gap in 5G energy management: the trade-off between collaborative efficiency and data privacy.

Economic Impact: It demonstrates that MNOs can significantly lower OPEX by pooling resources, turning high upfront CAPEX into long-term savings.
Scalability: The benefits of sharing increase with network density, making this solution particularly relevant for dense urban 5G deployments.
Sustainability: By optimizing the use of renewables and shared storage, the framework supports the transition to greener mobile networks.
Practicality: The use of Federated Learning provides a realistic pathway for competitors (MNOs) to cooperate on infrastructure without compromising proprietary data, potentially accelerating the adoption of shared energy ecosystems.

In summary, the "Collective Grid" framework proves that privacy-preserving AI combined with mathematical optimization can transform mobile network energy management from a siloed, costly operation into a collaborative, efficient, and sustainable system.