The global shift toward urbanization has necessitated advanced solutions to manage scarce resources. Energy usage, in particular, has become a critical focus area, as buildings account for approximately 40% of total global energy consumption and 36% of CO₂ emissions. IoT-enabled EMS in smart cities is essential to balance energy demand, reduce wastage, and promote the integration of renewable energy sources. The motivation for this research stems from three interconnected observations:

I. Scale mismatch: Centralized EMS architectures cannot efficiently process the terabyte-scale data streams generated by modern smart city deployments.

II. Security imperative: Energy infrastructure represents critical national infrastructure, making it a prime target for sophisticated cyber threats.

III. Economic viability: Cost-effective deployment strategies are essential for ensuring equitable access to smart energy solutions across diverse urban contexts.

This paper aims to: 1) critically review existing literature on EMS in smart city IoT networks; 2) identify and formalize major technical and operational challenges; 3) propose a novel distributed framework integrating AI and blockchain technologies; and 4) provide quantitative evidence of performance improvements through simulation.

The remainder of this paper is structured as follows. Section 2 presents a comprehensive literature review. Section 3 formalizes the challenges in current EMS architectures. Section 4 introduces the proposed distributed AI-blockchain framework. Section 5 develops the mathematical model for energy optimization. Section 6 presents simulation results and comparative analysis. Section 7 discusses implications and limitations. Section 8 concludes the paper and outlines future research directions.

Research into energy management in IoT-enabled smart cities has increased over recent years, primarily focused on strategies for optimising power consumption across different urban sectors [1]. Technologies such as smart grids, predictive algorithms, and machine learning are commonly applied.

EMS architectures in IoT-based systems range from centralised models to decentralised, peer-to-peer networks. Current protocols, like MQTT and CoAP, enable efficient communication but vary in energy consumption based on network design [2].

Current communication protocols, including MQTT (Message Queuing Telemetry Transport) and CoAP (Constrained Application Protocol), enable efficient device-to-device communication but vary significantly in energy consumption based on network design and deployment context. MQTT, with its publish-subscribe model, is particularly suited for high-throughput scenarios, while CoAP excels in constrained environments with limited bandwidth and power.

AI and machine learning play a pivotal role in predictive maintenance and demand response systems in EMS. Studies by Carrera et al. [3] and Jiang et al. [4] demonstrate that predictive algorithms can significantly enhance energy efficiency by anticipating usage patterns. Alsamhi et al. [5] applied deep reinforcement learning to energy management in smart city IoT networks, achieving a 22% reduction in peak demand compared to rule-based controllers.

Park et al. [6] developed adaptive energy management strategies for smart buildings using Recurrent Neural Networks (RNNs), showing that sequence modeling is particularly effective for capturing temporal dependencies in energy consumption patterns. Arévalo and Jurado [7] extended this work by integrating AI with IoT for optimizing smart grid operations, demonstrating that ensemble methods combining multiple prediction models outperform single-model approaches.

Integrating blockchain technology with IoT for energy management has emerged as a promising research direction. Reyna et al. [8] provided a comprehensive analysis of blockchain-IoT integration challenges and opportunities, identifying scalability and consensus efficiency as primary concerns. Tariq et al. [9] demonstrated the applicability of blockchain to secure IoT-based healthcare systems, establishing principles transferable to energy management.

For energy-specific applications, blockchain enables transparent peer-to-peer energy trading, immutable audit trails for regulatory compliance, and decentralized identity management for device authentication. Yang and Wang [10] proposed privacy-preserving energy management using blockchain, addressing the tension between transparency and confidentiality through zero-knowledge proofs.

Despite significant advancements, critical gaps persist in the literature. First, few studies address the integration of heterogeneous IoT devices with diverse communication protocols, making interoperability a persistent challenge. Second, the computational overhead of blockchain consensus mechanisms in resource-constrained IoT environments remains insufficiently characterized. Third, existing frameworks rarely provide formal mathematical guarantees for convergence and optimality in dynamic energy optimization problems. Fourth, the economic dimensions of large-scale EMS deployment—including total cost of ownership and return on investment—are underexplored. This paper addresses these gaps through an integrated framework combining lightweight consensus, formal optimization theory, and practical deployment considerations.

The expansion of IoT devices in urban areas strains current EMS infrastructure, creating a need for scalable systems capable of handling large-scale data. Consider a mid-sized smart city with connected devices, each generating data at rate ri (bits per second). The aggregate data rate is given by:

R = ∑ r . (1)

total i

For a city with N = 10 devices and average ri = 1kbps, Rtotal = 1 Gbps. Centralized processing of this data stream requires substantial computational infrastructure and introduces latency proportional to network diameter. The scalability challenge can be formalized as maintaining acceptable latency L ≤ Lmax and throughput T ≥ Tminas N → ∞.

Diverse IoT devices in smart cities often employ incompatible communication protocols, data formats, and semantic models, making seamless energy management challenging [11]. Interoperability can be quantified using the semantic compatibility metric:

M shared

1 ∣ Sj ∣

I = ∑ . (2)

M total ∣ Sj ∣

shared total

where Sj represents the set of shared semantics between system j and the EMS, and Sj is the total semantic space of system j. Values of I approaching 1 indicate high interoperability.

Exchanging energy data between IoT devices and EMS exposes the network to cyber threats, requiring robust encryption and privacy-preserving mechanisms. The security challenge can be characterized by the attack surface A, defined as:

A = ∑ v ⋅ e . (3)

where vk is the vulnerability score of component k and ek is its exposure level. Minimizing A while maintaining system functionality represents a multi-objective optimization problem.

Balancing operational cost with energy efficiency is critical. Most EMS setups prioritize energy savings but can be expensive to implement, limiting widespread adoption. The Total Cost of Ownership (TCO) over a planning horizon T can be expressed as:

Copex(t) + Cenergy(t) − Ssavings(t)

TCO = C + ∑ . (4)

capex t

(1 + δ)

where δ is the discount rate, Ccapexis capital expenditure, Copex(t)is operational expenditure, Cenergy(t)is energy cost, and Ssavings(t) represents energy savings.

We propose an energy optimization framework using AI-based predictive analytics to adaptively manage power consumption in real time. The framework leverages deep learning algorithms—specifically, a hybrid CNN-LSTM architecture, to predict demand surges and dynamically adjust supply. The prediction model takes as input:

I. Historical energy consumption sequences {Et−1, Et−2, . , . . Et−n}.

II. Weather forecasts {Wt, Wt+1, . , . . Wt+h}.

III. Calendar features (day of week, holidays, special events).

IV. Real-time IoT sensor data.

The hybrid architecture combines convolutional layers for local feature extraction with LSTM layers for temporal dependency modeling, producing demand predictions with a horizon of h time steps.

To address interoperability and security, we propose a permissioned blockchain system based on a Practical Byzantine Fault Tolerance (PBFT) consensus variant optimized for IoT environments. The blockchain serves three primary functions:

I. Decentralized identity management: Each device receives a cryptographic identity, enabling trustless authentication.

II. Immutable audit trails: All energy transactions are recorded transparently for regulatory compliance.

III. Smart contract-based coordination: Automated demand-response programs execute via self-enforcing smart contracts.

The lightweight consensus mechanism reduces computational overhead by 67% compared to proof-of-work while maintaining Byzantine fault tolerance for up to f faulty nodes in a network of 3f + 1 participants.

A distributed EMS architecture is proposed to improve scalability. By decentralizing energy management, each city sector (district, neighborhood, or building cluster) manages its own energy demand and supply while coordinating with a central system for aggregated data analysis. The architecture comprises four layers:

I. Device layer: IoT sensors and actuators.

II. Edge layer: Local processing nodes for real-time control.

III. Fog layer: Sector-level coordination and blockchain nodes.

IV. Cloud layer: Global optimization and long-term analytics.

This hierarchical structure reduces communication latency and enables graceful degradation during network partitions.

To lower implementation costs, we recommend a phased approach where EMS functions are gradually incorporated based on priority areas. This strategy allows cities to budget effectively while still benefiting from EMS improvements. The phasing strategy prioritizes:

Phase 1. (Months 1-6): Critical infrastructure monitoring (hospitals, emergency services).

Phase 2. (Months 7-18): High-consumption sector integration (commercial buildings, industrial zones). Phase 3. (Months 19-36): Residential and transportation integration.

Phase 4. (Months 37-48): Full ecosystem optimization and peer-to-peer energy trading.

The total energy consumption in a smart city network over a time horizon T is modeled as:

E = ∑ ∫ P (t) ⋅ η (t) dt. (5)

total i i

where:

I. Etotal= total energy consumption (kWh).

II. Pi(t)= instantaneous power demand of device i(kW).

III. ηi(t)= efficiency factor of device iat time t(dimensionless, 0 < ηi ≤ 1).

IV. N= total number of devices.

For discrete-time implementation with sampling interval Δt, this becomes:

E = ∑ ∑ P [k] ⋅ η [k] ⋅ Δt, (6)

total i i

where K = T/Δt is the number of time steps.

For load balancing across sectors, we propose the following optimization function that minimizes the variance of energy usage across sectors:

min J = ∑ (E − E ̅) . (7)

{Ej} balance j

Subject to:

∑ Ej = Etotal, (8)

min max

Ej ≤ Ej ≤ Ej , ∀j ∈ {1, . , . . M}. (9)

where:

I. Ej= energy usage in sector j.

II. E ̅ = ∑ E = average energy usage across sectors.

III. M= number of sectors.

min max

IV. Ej , Ej = minimum and maximum energy bounds for sector j.

The Lagrangian for this constrained optimization problem is:

M M M 2 max

ℒ(E , λ, μ , ν ) = ∑ ( E − ̅ E) + λ (∑ E − E ) + ∑ μ (E − E )

j j j j j total j j j j=1 j=1 j=1

(10)

min

+ ∑ νj (Ej − Ej).

Applying the Karush-Kuhn-Tucker (KKT) conditions yields the optimal solution:

λ μj νj

∗ ̅ (11)

Ej = E − − + .

2 2 2

The demand response flexibility of the system can be modeled using an elasticity matrix. Let direpresent the demand of consumer i and represent the electricity price. The price elasticity of demand is:

∂di/di

ϵij = . (12)

∂pj/pj

The total demand response ΔDto a price signal Δpis:

di

ΔD = ∑ ∑ ϵ ⋅ ⋅ Δp . (13)

ij j

pj

The energy demand prediction problem can be formulated as minimizing the regularized loss function:

min ℒ(θ) = ∑ (E ̂[k] − E[k]) + λ ∥ θ ∥ . (14)

θ reg 2

K k=1 where:

I. E ̂[k]= predicted energy demand at time k.

II. E[k]= actual energy demand at time k.

III. θ= model parameters.

IV. λreg= regularization coefficient.

For the LSTM component, the hidden state update equations are:

f = σ(W ⋅ [h , x ] + b ), (15)

t f t−1 t f

i = σ(W ⋅ [h , x ] + b ), (16)

t i t−1 t i

C ̃ = tanh(W ⋅ [h , x ] + b ), (17)

t C t−1 t C

C = f ⊙ C + i ⊙ C ̃ , (18)

t t t−1 t t

o = σ(W ⋅ [h , x ] + b ), (19)

t o t−1 t o

h = o ⊙ tanh(C ). (20)

t t t

where σis the sigmoid function, ⊙ denotes element-wise multiplication, and ft, it, ot represent forget, input, and output gates respectively.

For the PBFT-based consensus, the probability of successful consensus in the presence of faulty nodes among ntotal nodes is:

k n−k

P = ∑ ( ) p (1 − p) . (21)

consensus

k=0 k where p is the probability of a node behaving honestly. The system remains secure as long as f < n/3. The communication complexity of PBFT is O(n ), which can be reduced to O(n log n)using hierarchical consensus groups, making it suitable for large-scale IoT deployments.

The complete EMS optimization problem can be formulated as a multi-objective optimization:

min F(x) = [f , (x), f , (x)f (x)f (x)] . (22)

x 1 2 3 4

where:

I. f1(x)= total energy cost.

II. f2(x)= carbon footprint.

III. f3(x)= peak load demand.

IV. f4(x)= system vulnerability.

The Pareto-optimal front is obtained using the weighted sum method:

min ∑ w f (x). (23)

x i i

subject to ∑i=1 wi = 1, wi ≥ 0, and operational constraints.

The proposed framework was simulated using a synthetic smart city dataset comprising 10,000 IoT devices across 20 sectors over a 90-day period. The dataset was generated based on statistical distributions derived from real-world smart city deployments. Key parameters are summarized in Table 2.

Parameter Value Unit Number of devices (N) 10,000 -

Number of sectors (M) 20 -

Simulation period 90 days

Sampling interval (Δt) 15 minutes

Training/validation split 80/20 %

LSTM hidden units 128 -

Learning rate 0.001 -

Batch size 64 -

Blockchain block time 5 seconds

The simulation compared three architectures:

I. Centralized EMS (baseline).

II. Distributed EMS without AI (intermediate).

III. Proposed AI-Blockchain EMS (full framework).

Table 3 presents the demand prediction accuracy metrics for the AI component.

Metric Centralized Distributed (no AI) Proposed MAE (kWh) 12.4 10.8 6.2 RMSE (kWh) 18.7 15.3 8.9 MAPE (%) 9.8 8.4 4.7 R² 0.82 0.86 0.94

The proposed framework achieves a 49.5% reduction in RMSE compared to the centralized baseline, demonstrating the effectiveness of the hybrid CNN-LSTM architecture.

Table 4 summarizes the energy efficiency improvements observed across all scenarios.

Metric Centralized Distributed (no AI) Proposed Improvement Peak load (MW) 45.2 41.8 32.8 27.4% Daily energy (MWh) 892 847 614 31.2% Load factor 0.72 0.76 0.88 22.2% Renewable utilization 34% 41% 58% 70.6%

The proposed framework achieves a 27.4% reduction in peak load demand and a 31.2% improvement in overall energy utilization. The load factor improvement from 0.72 to 0.88 indicates significantly better demand-supply matching.

Table 5 presents the security analysis results under simulated attack scenarios.

Attack Type Centralized (Success Rate) Proposed (Success Rate) DDoS 68% 12% Man-in-the-middle 45% 8% Data tampering 72% 3% Sybil attack 55% 9% Replay attack 38% 5%

The blockchain-based architecture reduces successful attack rates by 82-96% across all attack types, demonstrating substantial security improvements.

Figure 1 (described) illustrates the latency scaling behavior. The proposed distributed architecture maintains sub-second latency for up to 100,000 devices, while the centralized architecture exceeds 5 seconds at 20,000 devices. The communication overhead scales as:

Lproposed = α log(N) + β. (24)

where α = 0.12 ms and β = 8.3 ms, compared to the linear scaling of centralized systems.

Over a 5-year horizon with a discount rate of 5%, the total cost of ownership analysis reveals:

I. Centralized EMS: $12.4M (capex) + $8.7M (opex) = $21.1M.

II. Proposed EMS: $15.8M (capex) + $4.2M (opex) = $20.0M.

Despite higher initial capital expenditure, the proposed framework achieves a 51.7% reduction in operational expenditure, resulting in net savings of $1.1M over five years and a break-even period of 3.2 years.

The results demonstrate that the proposed AI-blockchain EMS framework offers substantial improvements in energy efficiency, security, and scalability compared to traditional centralized approaches. The 31.2% reduction in daily energy consumption translates to significant environmental benefits, potentially reducing urban carbon footprints by 25-30% when scaled to city-level deployments.

Several limitations should be acknowledged. First, the simulation relies on synthetic data; real-world validation across diverse urban contexts is necessary. Second, the blockchain consensus mechanism, while optimized, still introduces latency that may be problematic for ultra-low-latency applications. Third, the framework assumes reliable network connectivity, which may not hold in all urban environments.

Compared to the framework proposed by Nikpour et al. [12], which achieved 18-25% energy reduction, our approach demonstrates superior performance through the integration of AI-driven prediction and blockchain-based security. The work of Du et al. [13] achieved 22% peak demand reduction; our 27.4% represents a meaningful improvement attributable to the distributed architecture and hybrid prediction model.