Experimental Demonstration of Cost-Effective Real-Time Risk Prediction and Energy Management System for Fog-Cloud-IoT Architecture in Smart City Applications

3.1 IoT fog network

The implementation of numerous sensing and actuation modules in smart homes necessitates maintaining connectivity with the IoT cloud, which can be resource-intensive. An IoT fog node addresses this by acting as an edge device that connects to both the local server and the cloud, ensuring data is logged locally and then reported. This setup offers benefits such as reduced latency, better scalability, and support for a range of devices. The fog node operates in online and offline modes: offline mode allows local network appliances to use without internet, synchronizing appliance states with the cloud upon reconnection. User preferences are stored on both the fog node and the cloud for direct appliance control. The node uses the MQTT protocol for communication with devices, hosts a local intelligence module for user-defined functions, and has the described service that facilitates data forwarding between different application programming interfaces (APIs) and includes features for device identification and user authentication. It utilizes Transmission Control Protocol (TCP) and Secure Socket Layer (SSL) for secure connections to cloud services. The architecture combines functionalities such as Secure Shell (SSH), Secure File Transfer Protocol (SFTP), alongside Web and Android APIs for configuring fog nodes. It features multiple security layers, including a traffic management firewall, Suricata IDS/IPS, and MISP, a platform for real-time threat intelligence exchange. MISP, integrated with Raspberry Pi, serves as the central hub for collecting Indicators of Compromise (IoCs) from various open-source sources. The proposed fog node, designed for smart home environments, utilizes a Raspberry Pi 4 featuring a quad-core processor and 8 GB of RAM.

This node utilizes services like SSH and SFTP for configuring smart homes while facing limitations in Wi-Fi range and request concurrency. To mitigate Wi-Fi issues, aesthetic reflectors and multithreading are introduced, but too much threading can harm performance. The architecture includes various security tools such as a firewall, Suricata IDS/IPS, and the MISP open-source threat intelligence platform for enhanced cybersecurity measures specific to smart homes. A custom Python script coordinates Suricata's interactions with MISP for real-time updating of threat intelligence. In case of significant alerts, Suricata sends notifications to MISP for widespread intelligence sharing and routes them for further analysis through Cohere’s AI API. Users are promptly informed via email alerts. The system functions on a Raspberry Pi as an IoT device in the smart home, while MISP connects to open-source threat feeds on a virtual machine, optimizing performance under hardware constraints.

3.2 IoT cloud network

The intelligent home architecture utilizes integrated IoT cloud technology designed for enhanced interactions among fog nodes, users, and developers. It incorporates a TCP/IP stack for connection-oriented services secured by SSL, along with a data dissemination layer for organizing appliance and sensor data before storage in the database. A copy of the database is maintained on a separate VM in the cloud, allowing for machine learning and data analytics while maintaining data integrity. Service-side requests for device access are managed by a relaying service, and customized Android and Web APIs facilitate user control over smart home appliances. A monitoring module for alert generation is also included, and additional protocols like FTP, SFTP, and SSH are utilized for secure communication.

3.3 Data transformation

The integration of various wireless protocols in a multi-layer WSN requires effective data exchange among devices using differing communication standards. Each protocol, such as Z-Wave, Zigbee, or Wi-Fi, uses distinct data formats, necessitating data conversions to ensure uniformity and functionality across the system. Methods for data transformation are crucial to accommodate these varying formats, which may include adjusting data fields or modifying packet structures to pertinent protocols. Approaches such as using JSON or XML can facilitate the data conversion process while maintaining data integrity across network layers, minimizing the need for specialized algorithms. Ensuring compatibility also requires structural alignment between the differing protocols. Aligned data formatting is crucial during data transformation to ensure compatibility with the receiving device. Minimizing processing time is a critical aspect, as delays can cause discrepancies and risk information loss. Thus, the development of effective transformation algorithms and efficient protocols is vital.

3.4 Energy management system

Forecast-Based Energy Prediction leverages predictive models and historical data to project future energy consumption or generation using various forecasting techniques like time-series analysis and machine learning. Accurate energy estimates allow for improved organization of energy generation, optimized supply chain management, cost reductions, and enhanced grid stability while diminishing issues such as power outages and low voltage outages in a smarter energy system. A proposed multi-hop forecast-based energy distribution model is described, aimed at predicting energy needs on multiple levels, both for individual households over different time frames and for groups of sub-grid level at smart homes. It manages overall energy demand, reduces waste, schedules energy usage, detects losses, and ultimately contributes to efficient energy supply management within smart areas. The proposed energy forecasting model employs an ensemble learning approach consisting of two stages: individual predictions and ensemble synthesis. Initially, three distinct models predict hourly energy demands, with outputs further processed by an ANN to enhance accuracy. The model focuses on devising mathematical formulations and algorithms for optimizing energy use in Home Energy Management Systems (HEMS), aiming to minimize total energy consumption and costs in smart urban settings. Key factors influencing energy optimization include power balance, the diversity of devices, energy sources, and temporal consumption patterns. Appliance sets within smart homes are represented as $T=\{a_1,a_2,a_3..\dots ,a_n\},$ where $a_1,a_2,a_3,..\dots ,a_{n,}$ the total energy consumed can be represented as ${\mathit{Tot}}_{\mathit{\text{ENERGY}}}^{SH}$ and specific consumption by the $j^{\mathit{th}}$ appliance as ${\mathit{\text{Energy}}}_{a_j}.$

These methodologies are applicable in diverse areas such as business, finance, supply chain management, meteorology, and energy. Time-series analysis, featuring ARMA and ARIMA models, is particularly important in evaluating energy consumption in smart cities, with ARMA models integrating Auto Regressive (AR) and Moving Average (MA) components for future value projections by referencing prior observations in relation to their current counterparts.

The MA part of the ARMA model analyzes the relationship between current observations and past residual errors from a moving average to forecast future results. Here, c signifies a constant, ${\mathrm{ß}}_1,{\mathrm{ß}}_2,\dots ,{\mathrm{ß}}_n$ represent coefficients, and $e_t$ denotes the error term.²

ARIMA models enhance ARMA by integrating an “I” component to manage non-stationary time series data through differencing. They are characterized by three parameters: p (autoregressive order), d (necessity of differencing for stationarity), and q (moving average order). SARIMA expands upon this framework by incorporating seasonal elements into the ARIMA structure, making it applicable for forecasting seasonal patterns and various forecasting scenarios. Where c represents a constant and ${\mathrm{ß}}_1,{\mathrm{ß}}_2,\dots ,{\mathrm{ß}}_n$ denote the coefficients.²

The ensemble model integrates three forecasting methods: ARMA, ARIMA, and SARIMA, to produce weekly forecasts. The generated weekly forecast data is then used as input for a simple ANN. Furthermore, a demand-based distribution phase is developed to allocate energy from the substation to smart homes, considering each home's specific needs and the energy supply from the grid substation. In general ANN structure consists of input layer, hidden layers and output layer. It performs certain actions such as forward propagation, loss function analysis, backward propagation and optimization is required. ANN structure is mainly used here to handle the data collected from the sensors like temperature, humidity and energy meter and so on.

3.5 Risk prediction

Risk prediction is the final stage of the proposed methodology. This phase focuses on quantitative measurement and risk analysis after identifying assets, threats, and vulnerabilities. It quantifies the risk from a specific threat to an asset, denoted as $R_{{\mathit{\text{threat}}}_{\mathit{ij}}}$ , by assessing both the likelihood of the threat and its potential impact on the asset, following a specific calculation method.

$R_{{\mathit{\text{threat}}}_{\mathit{ij}}}$ reflects the risk to asset i from threat j, $L_{{\mathit{\text{threat}}}_{\mathit{ij}}}$ assesses the probability of the occurrence of that threat, and $I_{{\mathit{\text{threat}}}_{\mathit{ij}}}$ evaluates the potential impact on the asset if the threat occurs. The total risk of an asset, $R_{\mathit{\text{asse}}{t}_i}$ , is obtained by aggregating the individual risks from all identified threats related to that asset. An equation is provided for determining this risk.

The risk associated with asset i is represented as $R_{\mathit{\text{asse}}{t}_i}$ , while $M_i$ signifies the total number of threats to that asset. The risk contribution of each specific threat j to asset i is denoted as $R_{{\mathit{\text{threat}}}_{\mathit{ij}}}$ . To standardize the risk evaluation for an asset, the maximum potential risk as influenced by all threats operating at their maximal impact has been evaluated with a mathematical expression.¹⁸

The asset risk indicates that $R_{\mathit{max}}\mathit{\text{asset}}$ is the utmost potential risk, with N representing the maximum number of threats and $R_{\mathit{max}}$ threat denoting the greatest risk assigned to any individual threat. It introduces a standardization equation that scales asset risk to a range of 0 to 100.

A smart home, denoted as $R_{\mathit{\text{normalized}}}{\mathit{\text{asset}}}_i$ . Lower normalized risk values indicate reduced risk, whereas values near 100 signify high risk. The overall risk of the smart home, represented as $R_{\mathit{\text{SmartHome}}}$ , is obtained by combining the normalized risks of all assets. An equation is referenced for determining this total risk.

$R_{\mathit{\text{SmartHome}}}$ represents the overall risk associated with a smart home, while N indicates the total number of assets present in the home. The term Rnormalized asseti signifies the overall risk of each asset within the smart home context, and Rnormalized asseti denotes the normalized risk associated with asset i. Normalization involves calculating the identified maximum potential risk that arises when all assets are exposed to their peak risk levels, determined through a defined equation.

In smart homes, the total number of assets is represented as N, while $R_{max\mathit{\text{asset}}}$ denotes the maximum potential risk for any individual asset. Overall risk is assessed by standardizing these measurements on a scale from 0 to 100 using specific equations.¹⁸

$R_{\mathit{\text{normalizedsmartHome}}}$ (normalized risk), $R_{\mathit{\text{SmartHome}}}$ (risk level), and $R_{\mathit{\text{MaxsmartHome}}}$ (maximum potential risk). The normalized risk is expressed as a percentage scale where values between 0-25 are designated as low risk, 26-50 as medium, 51-75 as high, and 76-100 as critical. Additionally, it introduces a systematic algorithm designed for assessing risks associated with individual assets. This Algorithm 1 classifies assets, identifies associated threats, calculates their likelihood and impact using established methodologies, and ultimately assesses the total risk for each threat by multiplying the two reported scores.

The CREFCI methodology supports cybersecurity experts by providing specific steps to evaluate a smart home's vulnerability to risks. Cumulative risk assessments enable experts to identify and categorize risk levels associated with various assets as critical, high, medium, or low, allowing for the prioritization of mitigation strategies tailored to individual asset risks. If high or critical risks are identified, the expert can reassess the individual asset risks to effectively address the most significant threats. The CREFCI model creates an advanced smart home infrastructure by combining IoT fog and cloud networks with effective forecasting and risk management techniques. It enables low-latency communication, secure data handling, and compatibility between various devices. The system optimizes energy consumption, reduces waste, and enhances grid reliability through a multi-hop forecasting approach. Proactive security measures are implemented through risk prediction, while intrusion detection and prevention are strengthened by integrating Suricata IDS/IPS and MISP. Seamless data communication across wireless standards, along with offline functionality, ensures continuity during connectivity issues.

4.1 Energy prediction of CREFCI

An energy prediction, scoring 75% for the highest probability predictions, a 13% enhancement over current models, which performed at 62%. For second-highest predictions, CREFCI achieved 45% compared to 32% from existing models, and for third-highest probabilities, it doubled the performance from 6% to 12%. It integrates IoT fog-cloud architecture, energy forecasting methods like ARMA and ANN, and multi-hop distribution techniques, enhancing prediction accuracy and reducing data latency in Figure 2.

Figure 2. Energy prediction of CREFCI.

4.2 Mean absolute error of CREFCI

It is to measure the average magnitude of error among the given data and with the prediction. It clarifies the true observation about the considered information and its absolute differences. A MAE of 0.004, corresponding to a 50% enhancement from an existing MAE of 0.008. Similar improvements were observed in the second-highest probability event, where the MAE decreased from 0.0125 to 0.007 (a 44% increase in precision), and in the third event, with a reduction from 0.020 to 0.012 (a 40% enhancement) in Figure 3. These improvements are attributable to a multi-hop forecasting mechanism combined with ensemble learning models and an ANN. The data transformation protocols ensure better compatibility among IoT devices, assisting in consistent data collection while effectively managing energy demand fluctuations.

Figure 3. Mean absolute error of CREFCI.

4.3 Consumed energy calculation

The energy consumption for EMSHF and EAIISS was 1.0 kWh, while CRASH registered 1.7 kWh. CREFCI utilizes only 0.5 kWh (50%-70% lower). In Smart mode, EMSHF and EAIISS consumed 0.9 kWh and CRASH consumed 1.6 kWh, whereas CREFCI required 0.4 kWh, promoting a 55%-75% in Figure 4. Under the optimally functional Smart + PF-PEC mode, CREFCI's consumption fell to 0.3 kWh, marking up to an 80% gain compared to others. This underscores the effectiveness of CREFCI's multi-hop energy prediction prototype, as enabled by its IoT fog-cloud architecture to lower data processing times and enhance power distribution. CREFCI uniquely integrates anomaly detection to avoid overconsumption linkage that spatial static methodology, including EMSHF and EAIISS, typically misses.

Figure 4. Consumed energy calculation.

4.4 Average latency calculation

It is the total time taken to perform a proper communication in response and return manner. It is defined as that of the elapsed duration of the transmitted required and the response delivered. In latency from 0.7 seconds (in EMSHF) to 0.3 seconds, achieving a 40-60% enhancement over baseline models due to its integration of IoT fog-cloud technology in Figure 5. This technology processes data closer to the source with edge and fog computing, expediting real-time decision-making and minimizing cloud-associated delays. A multi-hop energy prediction mechanism that helps distribute computational tasks effectively, adaptive scheduling for energy-efficient handling of time-sensitive events, and predictive analytics through ARIMA-ANN models for proactive data processing.

Figure 5. Average latency calculation.

4.5 Energy conservation calculation

An energy efficiency of 38%, surpassing the EMSHF (35%), EAIISS (34%), and CRASH (36%) models by 5–10%. The multi-hop energy forecasting and real-time optimization algorithms, which adapt energy resource allocation to fluctuations in demand. CREFCI enhanced thermal comfort, raising the PMV index from 1.0 (EMSHF) and 1.2 (EAIISS) to 1.8 in Figure 6. This IoT fog-cloud integration is used to dynamically regulate environmental parameters. The competing models relied on more static energy allocation tactics, resulting in limited energy savings and comfort.

Figure 6. Energy conservation calculation.