EcoFL: Green Federated Learning & Energy Systems

• EcoFL is a research umbrella that merges energy-aware federated learning with environmental and cyber-physical energy systems, emphasizing both energy constraints and performance.
• It employs innovative methodologies such as energy/carbon-aware optimization, communication-efficient architectures, and standardized carbon accounting to balance model accuracy with resource limits.
• EcoFL applications span distributed energy management and wireless resource allocation, achieving significant reductions in energy consumption and carbon emissions in practical deployments.

EcoFL denotes a family of research directions at the intersection of federated learning, environmental sustainability, and cyber-physical energy systems. In recent literature, the term is used both for green federated learning—FL whose design elements and parameters are chosen to decrease energy consumption while maintaining competitive performance and training time—and for federated learning applied to eco-/energy systems, especially where privacy-preserving forecasting is coupled to distributed control or market optimization in electric grids (Thakur et al., 2024, Du et al., 2023). It is also used as the name of concrete system architectures, including an O-RAN-based multi-RAT framework for energy-efficient FL resource allocation (Salama et al., 29 Jul 2025).

1. Terminology and scope

The literature does not use EcoFL as a single canonical algorithm. Instead, the term appears in several closely related senses.

Usage of “EcoFL”	Representative work	Core focus
Green federated learning	"Green Federated Learning: A new era of Green Aware AI" (Thakur et al., 2024)	Energy/carbon-aware FL design
FL for eco-/energy systems	"Federated Learning Assisted Distributed Energy Optimization" (Du et al., 2023)	Forecasting plus distributed energy optimization in a TEC
Named wireless resource-allocation framework	"EcoFL: Resource Allocation for Energy-Efficient Federated Learning in Multi-RAT ORAN Networks" (Salama et al., 29 Jul 2025)	O-RAN control of RAT selection, power, and PRBs
Standardized measurement methodology	"Standardized Methods and Recommendations for Green Federated Learning" (Tapp et al., 30 Jan 2026)	Phase-aware energy and CO$_2$e accounting

This suggests that EcoFL functions less as a single protocol than as a research umbrella spanning objective design, systems optimization, measurement methodology, and domain-specific deployment. Across these variants, the common thread is that FL is not evaluated solely by statistical accuracy: communication cost, computation cost, latency, energy consumption, and increasingly CO$_2$e are elevated to first-class design criteria (Thakur et al., 2024, Tapp et al., 30 Jan 2026).

2. Core optimization criteria and system models

In the green-FL sense, EcoFL starts from the standard FL objective

$$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$

and then adds energy or carbon constraints around it (Thakur et al., 2024). The survey literature explicitly frames green FL as minimizing total energy consumption subject to accuracy and training-time requirements, or as a multi-objective trade-off between energy and performance (Thakur et al., 2024).

A recurring systems model decomposes FL cost into computation and communication:

$$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$

with communication energy tied to message size, bandwidth, SNR, and transmit power, and computation energy tied to local epochs, CPU cycles, and CPU frequency (Thakur et al., 2024). Carbon accounting then maps power or energy to emissions through grid carbon intensity:

$\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$

or its discrete approximation (Thakur et al., 2024).

Two variables are emphasized as strongly correlated with FL’s carbon footprint: training time to reach target accuracy and the number of participating clients per round (Thakur et al., 2024). This is a crucial EcoFL insight because it shifts attention from raw algorithmic novelty to orchestration variables such as local epochs, client selection, model size, bandwidth allocation, and aggregation frequency.

Architecturally, the survey distinguishes centralized server-client FL, cross-device FL, cross-silo FL, and hierarchical or edge-assisted FL, and links each to different energy profiles (Thakur et al., 2024). Cross-device FL is especially sensitive to battery limits and wireless communication cost, while cross-silo FL shifts emphasis toward data-center and WAN carbon cost. A plausible implication is that EcoFL is inherently architecture-dependent: the same learning rule can have very different environmental behavior depending on where computation happens and how model updates move through the system.

3. EcoFL in distributed energy systems

In energy-system applications, EcoFL refers to FL embedded directly into distributed energy management. A concrete example is a Transactive Energy Community (TEC) composed of 125 residential prosumers with rooftop PV, organized into community $a$ with 100 buildings and community $b$ with 25 buildings, and interacting with several Virtual Power Plants through a distributed market mechanism (Du et al., 2023).

The forecasting layer runs at building level. Each building trains two local LSTM-based neural networks—one for demand and one for PV generation—with identical architectures: 3 LSTM layers with ReLU activation, two Dropout(0.2) layers, and a dense output layer. Shared hyperparameters include local epochs $e=8$, batch size $b_s=4$, validation split $v_s=0.25$, past observations $_2$0, and forecast horizon $_2$1 at 15-minute resolution (Du et al., 2023). Net demand is formed by subtraction,

$_2$2

and community demand is the building-wise sum (Du et al., 2023).

The FL layer is centralized, server-based horizontal FL using FedAvg with equal averaging over sampled buildings. Community $_2$3 uses $_2$4 rounds with $_2$5 buildings per round, while community $_2$6 performs a transfer-style adaptation with $_2$7, $_2$8, starting from the global model trained on community $_2$9 (Du et al., 2023). Raw load and PV time series remain local; only model weights are exchanged.

The optimization layer is a distributed Consensus + Innovations scheme. Each agent—either a VPP or the TEC community manager—maintains a local copy of the energy price $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$0, exchanges it with neighbors, and updates it through a consensus term plus an innovation term driven by local power or community demand. The critical systems point is that the innovation term depends on the community net-demand forecast. Forecast error therefore propagates into the distributed optimization, affecting both convergence and price accuracy (Du et al., 2023).

The empirical link is strong. In community $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$1, August demand RMSE is $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$2 for FL-forecasted demand and $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$3 for local-model forecasting, while the total daily energy price difference is $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$4 for FLF and $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$5 for LMF (Du et al., 2023). In community $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$6, April price difference falls from $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$7 under LMF to $$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$8 under FLF (Du et al., 2023). These results make EcoFL concrete in the energy-systems sense: FL is not merely a privacy layer for forecasting, but a component that measurably improves downstream distributed dispatch and price formation.

4. Network-aware and communication-efficient EcoFL architectures

A second major branch of EcoFL treats wireless and edge-network orchestration as the primary locus of energy efficiency. The named EcoFL framework for multi-RAT O-RAN networks combines a reinforcement-learning rApp in the non-RT RIC for RAT and power-plan selection with a CNN-based xApp in the near-RT RIC for PRB-allocation policy selection (Salama et al., 29 Jul 2025). The simulated deployment includes 1 LTE macro eNB at 10 MHz and 800 MHz, 1 5G NR micro gNB at 20 MHz and 3.5 GHz, and 50 clients. Voice traffic requires 0.1 Mbps and 100 ms delay; eMBB traffic, including FL, requires 10 Mbps and 80 ms delay (Salama et al., 29 Jul 2025).

The RL controller optimizes a reward

$$\min_{\omega} L(\omega), \qquad L(\omega)=\frac{1}{|C|}\sum_{c\in C} L_c(\omega),$$9

balancing energy efficiency, throughput, and latency (Salama et al., 29 Jul 2025). The xApp selects one of four predefined PRB policies: equal allocation, voice priority, eMBB priority, or dedicated reservation (Salama et al., 29 Jul 2025). Under this design, EcoFL reports average power $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$0 W versus $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$1 W for the baseline, energy efficiency $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$2 versus $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$3, and CIFAR-10 accuracy $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$4 with loss $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$5 (Salama et al., 29 Jul 2025).

Communication-efficient FL mechanisms form a closely related EcoFL subliterature. EcoFed targets partitioned or split DNN training on resource-constrained devices by freezing a pre-trained device-side model, eliminating gradient transmission, and combining a replay buffer with 8-bit activation quantization; it reports up to $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$6 communication reduction relative to classic FL and up to $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$7 training acceleration (Wu et al., 2023). EcoLoRA addresses federated fine-tuning of LLMs through round-robin LoRA segment sharing, adaptive sparsification, and Golomb coding, reducing communication time by up to $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$8 and total training time by up to $$E_{\text{total}} = E_{\text{comp}} + E_{\text{comm}} = \sum_{r=1}^{R}\sum_{c \in C_r} \bigl(E_{c,r}^{\text{comp}} + E_{c,r}^{\text{tx}} + E_{c,r}^{\text{rx}}\bigr),$$9 without compromising performance (Liu et al., 20 May 2025). This suggests that communication-aware model design is one of the dominant operational mechanisms by which EcoFL is realized in practice.

5. Measurement, carbon accounting, and resource-aware extensions

As the field broadened, EcoFL increasingly came to include not only energy-aware algorithms but also standardized environmental measurement. A phase-aware carbon-accounting methodology built on NVFlare and CodeCarbon instruments client processes into explicit tasks—init, idle_time, round_k, and evaluate—and adds an estimated communication-emissions term based on transmitted model-update size and a configurable network energy-intensity model (Tapp et al., 30 Jan 2026). In a CIFAR-10 workload, controlled client-efficiency degradation increased total CO$\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$0e by $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$1 in the medium-efficiency case and $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$2 in the low-efficiency case relative to the high-efficiency baseline (Tapp et al., 30 Jan 2026). In retinal optic disk segmentation, replacing V100 with H100 shortened runtime from roughly $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$3 to $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$4 minutes but produced non-uniform changes in total energy and CO$\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$5e across sites, reinforcing the need for per-site and per-round reporting (Tapp et al., 30 Jan 2026).

Another extension treats data and node selection as EcoFL levers. A data-centric approach to green FL analyzes data volume, label accuracy, consistency, and completeness, then predicts the data volume required for a target accuracy and ranks nodes by a score that combines CO$\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$6 impact with quality metrics (Sabella et al., 23 Jul 2025). In the reported heterogeneous time-series settings, Node Selection achieves an average $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$7 reduction in carbon emissions relative to baseline, with reductions up to about $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$8 in some cases; Smart Reduction attains the best success rate, meeting the target accuracy in $\text{CO}_2\text{e} = \int P(t)\,\gamma(t)\,dt,$9 of runs (Sabella et al., 23 Jul 2025).

Renewable-aware orchestration pushes the concept further by distinguishing grid energy from green energy. GreenFLag uses a Soft Actor-Critic controller to jointly allocate CPU frequency, transmit power, and bandwidth while accounting for communication contention, solar and wind harvesting, and battery state. Its objective minimizes grid energy over the FL process rather than merely minimizing total energy per round (Panagea et al., 31 Mar 2026). Using Copernicus data and a CNN-based MNIST workload, GreenFLag reports a $a$0 average reduction in grid energy consumption relative to three baselines, while primarily relying on green power (Panagea et al., 31 Mar 2026). A plausible implication is that mature EcoFL systems will increasingly optimize source-aware energy consumption, not only aggregate joules.

6. Open problems and research directions

The EcoFL literature identifies several unresolved tensions. The survey literature emphasizes heterogeneity versus energy/accuracy trade-offs, communication efficiency versus accuracy, and the lack of standardized benchmarks and reporting practices for energy and CO$a$1e in FL, especially in IoT deployments (Thakur et al., 2024). The carbon-accounting literature adds that inconsistent measurement boundaries and missing per-phase reporting make cross-paper comparison unreliable (Tapp et al., 30 Jan 2026).

Domain-specific deployments expose further challenges. In TEC energy management, non-IID demand and PV behavior, concept drift, peak prediction, and security risks such as model poisoning remain open issues, and more integrated designs could optimize directly for downstream market or dispatch criteria rather than forecast error alone (Du et al., 2023). In O-RAN-based EcoFL, the RL reward and xApp policy labels are hand-designed, the signaling overhead of the control loops is not explicitly quantified, and evaluation is limited to simulation rather than over-the-air deployment (Salama et al., 29 Jul 2025).

Broader wireless variants reinforce these trends. UAV-assisted FL has been formulated as a joint optimization of UAV trajectory, user participation, power allocation, and data volume control under an explicit convergence-accuracy constraint, solved through alternating optimization and successive convex approximation (Fu et al., 5 Aug 2025). Decentralized FL with overlapped clustering and D2D communications has been proposed to minimize device energy while maintaining convergence subject to time constraints, using bridge devices for decentralized aggregation without a global aggregator (Al-Abiad et al., 2022). These lines suggest that future EcoFL will increasingly co-design learning, networking, mobility, and energy infrastructure rather than treating FL as an isolated optimization layer.

Taken together, the literature portrays EcoFL as a shift from accuracy-centric federated learning to resource- and environment-aware federated systems. In some settings the focus is privacy-preserving optimization of physical energy systems; in others it is communication-aware radio control, carbon accounting, data-centric selection, or renewable-powered orchestration. What unifies these uses is the requirement that FL be evaluated not only by model quality, but also by the energy, carbon, and systems cost of attaining it (Thakur et al., 2024, Tapp et al., 30 Jan 2026).