Cloud-Enabled Hybrid, Accurate and Robust Short-Term Electric Load Forecasting Framework for Smart Residential Buildings: Evaluation of Aggregate vs. Appliance-Level Forecasting

Highlights

What are the main findings?

• A cloud-enabled hybrid framework combining Seasonal ARIMAX, Random For-est, and LSTM achieves high accuracy for a short-term residential load forecasting.
• Appliance-level (AL) and aggregate forecasts (AF) are compared, showing that AF outperforms AL in reliability and accuracy.

What are the implications of the main findings?

• Multi-cloud parallel calculations with a two-out-of-three voting scheme enhances forecasting robustness by mitigating single-cloud failures.
• The proposed scalable solution supports smart-city energy management, demand response, and grid reliability.

Abstract

Accurate short-term load forecasting is vital for smart-city energy management, enabling real-time grid stability and sustainable demand response. This study introduces a cloud-enabled hybrid forecasting framework that integrates Seasonal Autoregressive Integrated Moving Average with Exogenous variables (SARIMAX), Random Forest (RF), and Long Short-Term Memory (LSTM) models, unified through a residual-correction mechanism to capture both linear seasonal and nonlinear temporal dynamics. The framework performs fine-grained 5 min forecasting at both appliance and aggregate levels, revealing that the aggregate forecast achieves higher stability and accuracy than the sum of appliance-level predictions. To ensure operational resilience, three independent hybrid models are deployed across distinct cloud platforms with a two-out-of-three voting scheme, that guarantees continuity if a single-cloud interruption occurs. Using a real residential dataset from a house in Summerlin, Las Vegas (2022), the proposed system achieved a Root Mean Squared Logarithmic Error (RMSLE) of 0.0431 for aggregated load prediction representing a 35% improvement over the next-best model (Random Forest) and maintained consistent prediction accuracy during simulated cloud outages. These results demonstrate that the proposed framework provides a scalable, fault-tolerant, and accurate energy forecasting.

1. Introduction

In the modern era, energy management is shaping the world: nearly every sector, whether directly or indirectly, depends on efficient energy utilization, from AI-driven data centers and the protection of the ozone layer to the preservation of ecosystems, advancements in construction, industry, medicine, and tourism.

Energy management in smart cities ensures a balance between production and demand, prevents resource waste, improves air quality, and safeguards the planet.

“Smart” turns appliances into measuring, labeled, and controllable assets, and that radically improves AL data collection and forecasting accuracy/efficiency.

“Smart” appliances offer multiple advantages for accuracy of data, and hence for forecasting. Smart appliances can natively stream instantaneous power, voltage, current, and cumulative energy, not only periodic kWh. Also, they secure machine-readable semantics, not raw points. Devices can expose states (e.g., “wash cycle”, “defrost”, “setpoint”) with consistent naming/units. That makes feature engineering automatic and portable across vendors/homes. Forecasting upgrades to “state-aware”, that is one can move from generic time-series to state-space/HSMM or sequence models (RNN/Transformer) that understand cycles and dwell times (e.g., washer, fridge, heat pump). Papers on bottom-up/AL STLF show markedly better fidelity when appliance characteristics and states are included [1].

If devices are truly smart, one can read the appliance directly. This means less dependencies on NILM: Non-Intrusive Load Monitoring (a.k.a. energy disaggregation): Authors of [2] show sampling, transient features, and deep learning, including federated learning push accuracy, however under limits of overlapping loads.

Accurate load forecasting, enables companies to optimally scheduling their activities and determining the appropriate amount of energy to generate in order to maintain stability.

1.1. Load Forecasting Categories and Challenges

Electrical load consumption (ELC) forecasting is typically categorized into: short-term load forecasting (STLF), mid-term load forecasting (MTLF), and long-term load forecasting (LTLF). This study focuses on developing an efficient framework for STLF in residential buildings and addresses several critical research questions (this STLF framework directly supports smart-city energy operations by using smart-meter and weather streams to forecast demand at fine granularity, enabling demand response, EV-charging coordination, and optimal dispatch of solar-plus-storage across neighborhoods):

• Are statistical methods (SM) more effective for load prediction in residential buildings compared to machine learning (ML) techniques?
• Does a hybrid approach that combines SM and ML provide superior results?
• Is it more accurate to predict the overall consumption of a residential building (RB), or to forecast the consumption of individual appliances and then aggregate the results?

Furthermore, this paper examines the impact of feature-selection techniques, methods for combining statistical and machine learning (ML) models, residual (remainder) prediction (RP), and cloud-based computing.

Table 1. Challenges of statistical load forecasting.

Parametric Statistical Methods Gap
Handling high variability and peaks
Non-stationarity of load data
Difficulty handling missing data
Inaccuracy at different time horizons
Difficulty incorporating external variables
Inefficiency in real-time forecasting
Computational constraints with large datasets
Autocorrelation and multicollinearity
Risk of overfitting in complex models
Capturing seasonality and periodicity

and

Table 2. Challenges of ML load forecasting.

Machine Learning Techniques Gap
Generalization challenges across different regions or time periods
Need for large, high-quality datasets
Hyperparameter tuning complexity
Overfitting with complex models
Limited effectiveness in incorporating domain knowledge
Difficulty interpreting model decisions (black-box nature)
Model retraining requirements for real-time data
Difficulty handling rare events or outliers
High computational requirements
Sensitivity to feature selection and engineering

summarizes challenges for statistical modeling and ML modeling.

1.2. Feature Selection

Feature selection techniques have long been employed in scientific research, and their effectiveness has been widely studied [3,4,5,6]. Each method has its distinct advantages. For instance, the ANOVA F-test assesses whether the mean of a continuous (numeric) feature differs significantly across the categories of the target, by comparing between-class to within-class variance [7], mutual information measures nonlinear dependencies between features and the target variable [8], information gain and other entropy-based scores are commonly applied in classification tasks [9], LASSO suppresses irrelevant features and thus performing implicit feature selection [10], and genetic algorithms offer evolutionary optimization for selecting feature subsets [11]. Research in [12] investigates the role of correlation matrices in prediction and showed that features with high correlation can enhance dataset performance. In [13], the authors evaluated the Chi-Square statistical test as a feature selection method, demonstrating its ability to assess whether features have a meaningful relationship with the target. Furthermore, the authors of [14] highlighted the strength of the Maximum Relevance Minimum Redundancy (MRMR) method in identifying features that are both relevant and non-redundant, thereby mitigating overfitting caused by redundant information.

This research proposes a novel framework that integrates a voting-based system for feature selection. By combining statistical methods with a hybrid approach, the framework leverages correlation analysis, feature independence, and redundancy avoidance to maximize predictive performance in residential load forecasting.

1.3. Statistical Methods (SM)

Statistical methods [15,16], machine learning (ML) techniques [17,18], and hybrid models [19,20] have continually evolved over the past decades, leading to significant advances in prediction accuracy.

For example, research in [21] demonstrated the effectiveness of the autoregressive integrated moving average (ARIMA) model for time-series forecasting, particularly in scenarios driven by temporal dependencies. However, ARIMA struggles to capture seasonality, a limitation that has been addressed by seasonal ARIMA (SARIMA) [22] and SARIMAX (seasonal ARIMA with exogenous regressors) [23]. Despite these improvements, statistical methods generally lack the scalability and flexibility of ML techniques when dealing with large datasets and complex nonlinear relationships.

1.4. Machine Learning (ML)

Machine learning methods, on the other hand, have proven highly powerful and are now widely applied across nearly all domains of science. In [24], the authors highlighted the accuracy of support vector machines (SVMs), which are supervised learning models used in both classification and regression tasks. Tree-based and sequence-based models have also emerged as dominant approaches in forecasting. Decision trees (DTs) [25] produce hierarchical decision rules by recursively splitting on feature values, offering interpretability but often suffering from overfitting. Random Forests (RFs) [26] overcome this drawback by aggregating multiple randomized trees, thereby reducing variance and improving robustness. Extreme Gradient Boosting (XGBoost) [27] further enhances ensemble learning by constructing trees sequentially to correct residual errors, incorporating gradient optimization, regularization, and efficient computation for superior predictive performance. In parallel, Long Short-Term Memory networks (LSTMs) [28] are designed to capture long-range temporal dependencies using gated memory cells, making them especially effective for sequential and time-series forecasting.

Nevertheless, a limitation of many ML-based approaches is their tendency to focus on identifying patterns without explicitly incorporating seasonality or linear temporal dynamics, areas where statistical models such as ARIMA, SARIMA, and SARIMAX excel. This contrast motivates development of hybrid frameworks, which combine the nonlinear modeling power of ML with trend- and seasonality-capturing strengths of statistical approaches. By integrating these complementary capabilities, hybrid systems achieve higher accuracy, reliability, and generalizability.

Authors in [29] present a short-term residential load forecasting framework that integrates evolutionary algorithms for neural architecture search with transfer learning to adapt models across multiple households. The method demonstrates superior performance over benchmark techniques, with an average accuracy gain of 3.17 percentage points compared to the state-of-the-art LSTM one-shot approach. Researchers in [30] introduced a hybrid forecasting framework for urban power load that identifies and quantifies abnormal variation periods via residual decomposition, augments scarce abnormal samples using WGAN-GP, and employs TimesNet to capture complex nonlinear patterns. The method boosts forecasting accuracy for abnormal load periods by about 38.9%. Although it incurs only a minor drop (2.65%) across all periods, it outperforms baseline deep-learning models. The authors of [31] proposed a hybrid deep-learning framework (GNN-ViGNet) enhanced by the use of IoT sensor data and novel preprocessing, feature engineering, and imbalance correction methods, coupled with a Coati–Northern Goshawk Optimization algorithm for real-time route optimization in EV fleets. Their model achieves 98.9% accuracy in charging load forecasting and reduces route distance to 511 km, significantly outperforming baseline techniques such as PSO and Firefly in both convergence speed and prediction error metrics. The authors of [32] introduced radian scaling, a new normalization method that transforms differences of consecutive electricity load values into radians to maintain normalized distributions under seasonal concept drift without retraining machine learning models. Their experiments using lightweight ANN models show radian scaling reduces RMSE dramatically from 158.63 to 43.38 and cuts training epochs by about 67%, statistically outperforming traditional normalization methods. The work in [33] develops a learning framework for forecasting electrical load in renewable energy communities with dynamic portfolios, synthesizing non-stationary, discontinuous, and NON-IID time series, adapting preprocessing and model input (past, present, future features) to achieve domain-invariant feature learning.

Their approach achieves forecasting accuracy comparable to that of centrally trained models, while outperforming single-source models by more than 18% in terms of MAE, and demonstrating strong transferability to unseen data. This research proposes a hybrid forecasting framework that integrates RF, LSTM, and SARIMAX models to address these challenges and enhance short-term load forecasting performance in residential buildings. As hybrid models have demonstrated high performance, this paper also investigates this approach and extends it by incorporating additional techniques to further enhance forecasting accuracy and robustness. Specifically, the study integrates efficient feature-selection methods to improve predictive performance and employs concurrent cloud computing together with a voting-based strategy to increase the model’s resilience and scalability.

1.5. The Contributions of the Presented Work Can Be Summarized as Follows

• Integration across paradigms: A unified forecasting architecture that synergistically combines Seasonal ARIMAX, Random Forest, and LSTM models through a residual-correction layer, effectively merging statistical trend modeling with nonlinear and sequential learning.
• Feature–model customization: Each sub-model employs an optimized feature selection strategy (Correlation Matrix, Chi-Square, or MRMR), enabling the framework to capture diverse predictive relationships with minimal redundancy.
• Multi-cloud redundancy and voting: A two-out-of-three cloud voting ensures continuous operation under individual cloud or model failures, an aspect rarely addressed in prior forecasting frameworks.
• Fine-grained temporal resolution: The framework is validated on 5 min interval residential load data, demonstrating that aggregate forecasts outperform the summation of appliance-level predictions (sum of predictions per appliances) in both stability and accuracy.
• Demonstrated accuracy and robustness: The system achieves a 35% reduction in RMSLE relative to the best single-model benchmark and maintains reliable performance under simulated cloud outages.

A detailed comparison between aggregate prediction and appliance-level forecasting in terms of accuracy and reliability. Altogether builds a scalable, fault-tolerant, and statistically grounded hybrid methodology for short-term load forecasting in smart residential environments.

The rest of the paper is organized as follows: Section 2: Data description and preparation, Section 3: Feature selection, Section 4: Methodology, Section 5: Discussion and comparison of the results, Section 6: Conclusions.

2. Methodology

This research proposes a novel method that integrates both statistical approaches and machine learning techniques, termed SSRLR (Sparse voting feature-selection technique, Seasonal Autoregressive Integrated Moving Average with Exogenous variables model, Random Forest, Long Short-Term Memory, and Remainder Prediction). This integration significantly enhances the overall predictive accuracy of the proposed framework. The machine learning component is responsible for capturing complex, non-linear patterns and intricate relationships within the data, while the statistical component simultaneously models seasonality effects and time-dependent correlations. By operating in parallel, these two components complement and reinforce each other: the statistical models address temporal regularities that machine learning alone might overlook, and the machine learning algorithms uncover high-dimensional interactions that are beyond the capacity of purely statistical methods. This mutual coverage enables the framework to extract the maximum amount of useful information from the data while mitigating the risk of overfitting. Furthermore, the synergy between the two paradigms ensures that weaknesses in one component are compensated for by the strengths of the other, thereby improving robustness and generalization performance across diverse scenarios.

Figure 1 illustrates the flow of the embedded procedures. After collecting data on electricity consumption and relevant weather variables, preprocessing is performed by applying the Holiday/Weekend factor and modeling the seasonal effects. Then, the process proceeds to the feature selection phase using three feature selection techniques: Chi-Square test, MRMR, and CM. These methods are applied to identify and retain the most informative features. By systematically selecting only the most relevant variables, this step extracts an optimal set of features, thereby improving generalization and reducing the risk of overfitting. Consequently, the feature selection phase plays a pivotal role in enhancing the robustness and accuracy of the forecasting framework.

Next, the process of primary prediction (PP) employs both machine learning and statistical methods, specifically, the Long Short-Term Memory (LSTM), the Seasonal Autoregressive Integrated Moving Average with Exogenous variables (SARIMAX) model, and the Random Forest (RF) algorithm. The SARIMAX model is utilized to capture and exploit the most informative seasonal effects and other time-dependent features, thereby maximizing the extraction of temporal patterns from the data. The RF method enables the exploration of diverse forecasting scenarios by leveraging its ensemble-based decision-making capability, which enhances robustness against overfitting. In parallel, the LSTM component contributes to the modeling of both short-term and long-term temporal dependencies, thus complementing the strengths of SARIMAX and RF. This hybrid integration substantially enhances the learning potential of the proposed method, enabling it to adapt effectively to complex and dynamic patterns present in the dataset.

A primary trend is first established to capture the general direction of the target variable. This trend enables the derivation of an additional and highly informative feature the Remainder Prediction (RP) in which deviations from the trend are referred to as remainders. Once the primary trend is obtained, the deviations of the actual values from the underlying forecasted trajectory are calculated. These deviations are then treated as a separate time series, which can be independently modeled to predict the future remainders. The final forecast, i.e., overall consumption, all individual appliance-level forecasting, and also their total consumption. that is subsequently obtained by summing the predicted primary consumption values and the predicted remainders. This decomposition-based strategy reduces forecasting error because it allows the model to focus on different components of data separately: the trend model specializes on capturing long-term patterns, while the remainder model focuses on short-term fluctuations and irregularities. By isolating these components, the approach mitigates the risk of error propagation that can occur when a single model attempts to fit both large-scale and fine-grained variations simultaneously.

This research further enhances the reliability, processing speed, and real-time computing capability of the proposed framework through the use of cloud computing. In this setup, three independent cloud instances operate in parallel. The final forecast is obtained through an averaging mechanism, while a majority-voting mechanism is employed to ensure that a correct prediction is selected. Specifically, if at least two of the three instances exhibit the same level of error (the difference is below a threshold), their output is accepted as correct. If the third instance shows a significantly larger error, it is excluded from the forecasting process and flagged for expert review and retuning. This redundancy significantly improves the accuracy and robustness, allowing the system to operate effectively even in the event of a single instance failure, and it minimizes the risk of erroneous outputs caused by transient faults or network disruptions. The adoption of this dual-stage forecasting methodology trend plus Remainder Prediction within a fault-tolerant cloud infrastructure offers several benefits. From a modeling perspective, separating long-term and short-term patterns allows for more specialized learning and reduces cumulative forecasting errors. From a system architecture perspective, the two-out-of-three cloud framework ensures high availability, operational continuity, and consistent performance under varying computational loads. Leveraging cloud computing brings additional advantages. The elastic scalability of cloud resources allows the system to handle sudden increases in data volume without sacrificing processing speed. Geographic distribution of cloud servers reduces latency, enabling near real-time decision-making. Furthermore, cloud platforms facilitate seamless integration of advanced analytics tools, automated backups, and security protocols, ensuring that forecasting models remain both robust and secure while being accessible from anywhere. This combination of advanced forecasting methodology and resilient cloud-based deployment provides a high-performance, dependable solution for real-world operational environments.

2.1. Long Short-Term Memory (LSTM) Networks

Long Short-Term Memory (LSTM) [34] networks, sketched in Figure 2, are a form of Recurrent Neural Network (RNN) specifically built to capture dependencies that span long intervals in sequential data. Unlike standard RNNs, which struggle to learn from distant inputs due to vanishing gradients during backpropagation, LSTMs introduce a dedicated memory cell that can preserve and carry information for many time steps. This design makes them especially powerful for applications like forecasting time series, understanding speech, and processing text.

At each time step, the memory cell’s content is carefully managed by three gates forget, input, and output. The forget gate decides which past information should be removed, the input gate controls what new information enters the cell, and the output gate determines what part of the memory contributes to the next hidden state. By adjusting these gates dynamically, an LSTM can flexibly balance learning short-term fluctuations and long-term trends within the data.

LSTMs have demonstrated strong performance in numerous areas. In natural language processing, they learn the flow of sentences to generate or translate text. In speech recognition, they interpret audio sequences to convert spoken words into text. They are also effective in spotting unusual patterns in time-series data for anomaly detection. Despite their benefits, LSTMs require substantial computational resources and their inherently sequential updates limit the ability to parallelize training compared to architectures like Convolutional Neural Networks (CNNs). Nevertheless, their capacity to handle both isolated observations and entire sequences makes them a cornerstone model in deep learning research involving temporal or ordered inputs.

2.2. Random Forest (RF)

The Random Forest [35] algorithm (see Figure 3) is an ensemble-based technique that leverages the simplicity of individual decision trees while dramatically improving their reliability. Instead of relying on a single tree, it grows dozens or even hundreds of them, each trained on a different “bootstrap” sample of the data (i.e., random draws with replacement). At each split in a tree, only a randomly chosen subset of the available features is considered, which encourages diversity among trees and helps prevent overfitting. When making a prediction, the forest either tallies a vote among all trees (for classification) or computes the mean of their outputs (for regression), yielding a consensus result that is typically more accurate and stable than any single tree.

One of Random Forest’s greatest assets is its ability to estimate the relative importance of each feature: by measuring how much the model’s error increases when a given variable’s values are permuted, insight is gained into which inputs truly drive predictions. This makes RF a valuable tool both for high-stakes applications such as detecting fraudulent transactions, forecasting electrical loads, or diagnosing diseases and for exploratory analysis, where understanding the data is as important as prediction itself. On the downside, training and storing a large collection of trees can demand considerable computation and memory, and making real-time predictions may be slower if the forest is very deep. Moreover, while each decision tree is easy to follow on its own, the combined forest becomes something of a “black box”, making it harder to explain individual decisions than with simpler, single-tree models. Despite these trade-offs, Random Forest remains a go to choose whenever robustness to noise, missing values, and outliers is required together with strong predictive power.

2.3. Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX)

The SARIMAX [36] model extends the familiar ARIMA approach by explicitly handling seasonal fluctuations and allowing for outside influences. It is ideal when the data repeat patterns at regular intervals such as monthly sales cycles or annual weather variations while also being affected by additional variables like temperature or economic indicators. In essence, SARIMAX merges three pieces:

• Non-seasonal ARIMA component:

Autoregressive (AR)p)):

This component models the current value of the series by referencing its own previous observations. The order parameter p designates how many past points are used to predict the present.

Integrated (I)d)):

To stabilize the series, the data are differenced d times. This process removes trends and helps ensure that statistical characteristics such as the mean and variance remain roughly constant throughout the sequence.

Moving Average (MA)q)):

Here, the forecast is formed by combining current and past forecast errors. The parameter q indicates how many of these lagged error terms are included in the smoothing equation.

Mathematically, this is written as:

$${\mathsf{\varphi }\left(\mathrm{B}\right) (1-\mathrm{B})}^d{y}_t=\mathsf{\theta }(\mathrm{B}) {\mathsf{ϵ}}_t$$

(1)

where

• ϕ(B): is the non-seasonal AR polynomial
• θ(B): is the non-seasonal MA polynomial
• B: is the backshift operator
• y_t: is the series at time t, and
• ${\mathsf{ϵ}}_t$: is a white-noise error term

2.

Seasonal Component (s)

To capture repeating patterns every “s” period, SARIMAX uses seasonal AR and MA polynomials of orders P and Q, together with seasonal differencing D. This portion is given by:

$$\mathsf{\Phi }\left({\mathrm{B}}^s\right){\left(1-{\mathrm{B}}^s\right)}^D y_t=\mathsf{\Theta }\left({\mathrm{B}}^s\right) {\mathsf{ϵ}}_t$$

(2)

where

• Φ(${\mathrm{B}}^s$): Seasonal autoregressive polynomial of order P.
• Θ(${\mathrm{B}}^s$): Seasonal moving average polynomial of order Q.
• S: Seasonality period
• D: Number of seasonal differences.

3.

Exogenous regressors ($\mathrm{X}$)

$$y_t=\mathsf{\beta }{\mathrm{X}}_{\mathrm{t}}+{\mathsf{ϵ}}_{\mathit{t}}$$

(3)

where

• ${\mathrm{X}}_{\mathrm{t}}$: External (exogenous) regressors at time t.
• $\mathsf{\beta }$: Coefficient vector for the exogenous variables.

Combining all components, the SARIMAX model can be represented as:

$${\mathsf{\varphi }\left(\mathrm{B}\right) (1-\mathrm{B}) }^d \mathsf{\Phi }\left({\mathrm{B}}^s\right){(1-{\mathrm{B}}^s)}^D y_{t=\theta }\left(\mathrm{B}\right) \mathsf{\Theta }\left({\mathrm{B}}^s\right){\mathsf{ϵ}}_t+ \mathsf{\beta }{\mathrm{X}}_{\mathrm{t}}$$

(4)

By combining these three elements non-seasonal dynamics, periodic behavior, and outside influences SARIMAX delivers a flexible framework for forecasting. It is widely used in examples like projecting retail sales with marketing spend as an exogenous input, estimating electricity demand driven by temperature and humidity, or any scenario where patterns repeat and external drivers matter.

2.4. SSRLR

The pseudocode of the proposed SSRLR is illustrated below (Algorithm 1):

SSRLR, a novel hybrid method that integrates statistical and machine learning techniques including a Sparse voting feature selection technique, Seasonal Autoregressive Integrated Moving Average Exogenous model, Random Forest, Long Short-Term Memory, and a Remainder Prediction step to leverage both linear trends and complex nonlinear relationships.

Algorithm 1. Pseudocode of proposed algorithm

Input: Historical load data, Weather features, Forecast horizon = 12 Step 1: Data Preprocessing   Normalize data using MinMaxScaler to range [0,1] Step 2: Feature Selection   Apply CM   Select relevant features → CM_Output   Use this for LSTM   Apply Chi-Square   Select relevant features → Chi-Square_Output   Use this for SARIMAX   Apply MRMR   Select relevant features → MRMR_Output   Use this for RF Step 3: Model Training and Prediction   Train SARIMAX model with seasonal and non-seasonal components   Predict next 12 points → SARIMAX_Output   Train Random Forest model using weather and historical data   Predict next 12 points → RF_Output   Train LSTM model for time-series forecasting   Predict next 12 points → LSTM_Output Step 4: Remainder Prediction (RP)   Perform additional adjustments if needed:   Use SARIMAX for LSTM on cloud 1   Use LSTM for RF on cloud 2   Use RF for SARIMAX on cloud 3 Step 5: Ensemble Averaging   SARIMAX Final Output (PP+RP) → cloud 1_Output   RF Final Output (PP+RP) → cloud 2_Output   LSTM Final Output (PP+RP) → cloud 3_Output Step 6: Ensemble Averaging   Ensemble_Output = (cloud 2_Output + cloud 2_Output + cloud 3_Output)/3 Step 7: Output Final Prediction   Return Ensemble_Output

The dataset is divided into 80% for training and 20% for testing, after which the following procedures are applied:

• The process begins with feature selection using the CM method, followed by primary forecasting with LSTM. A trend is then identified for the remainders, and SARIMAX is applied to forecast the subsequent remainders. The final output for this stage is obtained by summing the LSTM predictions with the forecasted remainders.
• Feature selection is performed using the MRMR method, followed by primary forecasting with Random Forest (RF). A trend is then identified for the remainders, and LSTM is applied to forecast the subsequent remainders. The final output for this stage is obtained by summing the RF predictions with the LSTM-forecasted remainders.
• Feature selection is performed using the Chi-Square method, followed by primary forecasting with SARIMAX. A trend is then identified for the remainders, and Random Forest (RF) is applied to forecast the subsequent remainders. The final output for this stage is obtained by summing the SARIMAX predictions with the RF-forecasted remainders.

Figure 4 presents the detailed schematic of the proposed method.

2.5. Two-Out-of-Three Multi-Cloud Voting Architecture

Single cloud has been used in load forecasting in previous research [37]. To enhance the robustness, accuracy, and real-time performance of the forecasting system, the three pipelines described in Section 2.4 are deployed across three independent cloud instances. As illustrated in Figure 5, each cloud instance executes one of the proposed hybrids forecasting strategies:

• Cloud 1: Implements the CM → LSTM → SARIMAX pipeline.
• Cloud 2: Implements the MRMR → RF → LSTM pipeline.
• Cloud 3: Implements the Chi-Square → SARIMAX → RF pipeline.

Figure 5. Proposed two-out of-three cloud-based system.

Figure 5. Proposed two-out of-three cloud-based system.

Each cloud processes the same input dataset in parallel, but with different feature-selection methods, primary forecasting models, and remainder forecasting techniques. This methodological diversity significantly reduces the risk of correlated model errors and increases the resilience of the overall system.

Preliminary Averaging: The forecasts from all three clouds are averaged to produce an initial consensus prediction.

Error-Based Majority Voting: The errors from each cloud are evaluated against the actual values. If at least two clouds yield similar and low error magnitudes, their results are accepted as correct. The third cloud, if it shows significantly higher error, is excluded from the current forecasting decision and flagged for expert review and model retuning.

This redundancy ensures that the system maintains robustness remaining operational and accurate even in the presence of network disruptions, computational errors, or degraded model performance in one instance. By leveraging agreement between at least two independent pipelines, the architecture minimizes the likelihood of passing erroneous forecasts to the end user or decision-making process. Figure 5 shows this cloud-based architecture.

The deployment of the two-out of-three architecture in a cloud environment provides several operational and strategic benefits:

• High Availability and Robustness: The redundancy of multiple clouds ensures continuous operation even if one instance experiences downtime or degraded performance.
• Elastic Scalability: Cloud resources can be dynamically allocated to handle varying computational demands, enabling real-time or near real-time processing for large-scale datasets.
• Parallel Computation: Hosting clouds in different regions improving forecast delivery speed and responsiveness.
• Security and Data Integrity: Cloud platforms provide encryption, controlled access, and automated backup systems, ensuring the safety and reliability of both input data and forecasting models.
• Cost Efficiency: Pay-as-you-go billing models allow for optimal use of resources without the overhead of maintaining extensive on-premises infrastructure.
• Seamless Integration and Maintenance: Cloud ecosystems support easy deployment of updates, integration of analytics tools, and automated monitoring of model performance with cheaper price compare to on- premise systems.
• Collaboration and Accessibility: Authorized users can securely access the forecasting system from anywhere, enabling collaborative work across geographically distributed teams.

In the suggested two-out-of-three cloud-based redundancy model, reliability is expressed in terms of R, the probability of each cloud (e.g., AWS) functioning correctly:

• R³ denotes the case where all three systems perform without error.
• 3R² (1 − R) corresponds to the case where two systems deliver correct outputs and one produces an error, which is automatically overruled by the majority decision.

The system failure occurs only if two or more clouds malfunction at the same time, a situation not covered by this formula.

By combining diverse forecasting models with a fault-tolerant two-out-of-three multi-cloud deployment, the proposed system achieves a high degree of accuracy, resilience, and computational efficiency. This design ensures that forecasts remain consistent, secure, and responsive under real-world operational conditions, even in the presence of hardware failures, software errors, or sudden spikes in data volume.

3. Data Description and Preparation

The dataset consists of power usage readings for a household in Summerlin, Las Vegas, Nevada, USA, sampled at five-minute intervals for 365 days of 2022. This dataset was officially collected from a residential customer of a local electric utility as part of its demand response program. Each measurement is given in kilowatts (kW) and reflects the consumption of the following household systems and appliances:

• Forced-air heating unit on the lower level
• Air conditioner on the lower level
• Refrigerator
• Air conditioner on the upper level of the house
• General (miscellaneous) household load
• Laundry machines (washer and dryer)
• Dishwasher and garbage disposal
• Microwave oven
• Forced-air heating unit on the upper level of the house
• Kitchen circuit (GFI) including the toaster

The dataset offers a full coverage with no gaps or missing entries, which makes it a solid foundation for any analysis. Electricity usage was logged every five minutes throughout 2022, giving a detailed picture that picks up even the smallest variations in consumption. Before feeding the data into the forecasting models, preprocessing is performed and relevant features are carefully selected. This step both smooths out irregularities and reduces the risk of overfitting. Figure 6 illustrates the data profile.

Estimating the Holiday/Weekend Adjustment Factor and Modeling the Seasonal Effect

Before feeding the data into the forecasting model, the dataset is preprocessed to adjust for unusual consumption patterns that occur during holidays and weekends, when usage often departs from typical weekday behavior. Social traditions, leisure activities, and economic regulations all play a role: some people travel or gather with friends, while others stay home using appliances or hosting events. Many people in Las Vegas work during the weekends, holidays, or change shifts frequently. In addition, manufacturing schedules may shift factories in some areas shut down for holidays, whereas others continue operating further altering the load profile. To quantify this effect, a Holiday Adjustment Factor is computed by first tagging each timestamp as either a “holiday/weekend” or a “weekday.” The data are then divided into two groups one representing holidays and weekends and the other representing regular weekdays and the mean energy consumption is calculated for each group. The adjustment factor is obtained by dividing the holiday/weekend average (1.89 kW) by the weekday average (1.94 kW), yielding a value of approximately 0.98. This value shows that, on average, consumption during holidays and weekends is about 98% of typical weekday use, and incorporating this factor into our model helps dampen the extra variability introduced by these irregular days, resulting in more accurate forecasts. (Holiday Coefficient = (Average Consumption on Holidays-Weekends)/(Average Consumption on Regular Days)).

Holidays included in 2022: 1 January (New Year’s Day, Saturday), 17 January (Martin Luther King Jr. Day, Monday), 21 February (Presidents Day, Monday), 30 May (Memorial Day, Monday), 4 July (Independence Day, Monday), 5 September (Labor Day, Monday), 10 October (Columbus Day, Monday), 11 November (Veterans Day, Friday), 24 November (Thanksgiving Day, Thursday), 25 December (Christmas Day, Sunday).

Next, each calendar month is standardized to a fixed length of 31 days. When a month contains fewer than 31 days, the readings from its final day are carried forward to fill the gap. This step creates a uniform time grid, enabling the model to learn genuine seasonal patterns without distortion from varying month lengths.

Seasonal effects are regular, repeating variations in energy consumption that occur over time hourly, daily, weekly, or monthly. By assigning each month a uniform 31-day span, artificial discontinuities in the data are eliminated, allowing the algorithm to detect genuine periodic patterns.

• Applying this procedure to our 2022 dataset produces:
• A total of 372 “normalized” days (12 months × 31 days)
• 372 days × 24 h = 8928 h
• At 5 min resolution: 8928 h × 12 readings per hour = 107,136 data points

To ensure temporal alignment across months, we standardized all monthly sequences to 31 days. This choice is consistent with prior studies in residential and urban load forecasting where month-length normalization is required. Importantly, the 31-day format preserves the complete set of real observations: months with fewer than 31 days are expanded using a small number of interpolated entries, while months with 31 days remain unchanged. In contrast, a 30-day standardization would require deleting one full day of real measurements from every 31-day month, which removes genuine load patterns and weakens the statistical richness of the dataset. Additionally, the 31-day format requires fewer temporal adjustments (7 days per year) compared to the 30-day approach (9 days per year), reducing distortion and maintaining closer fidelity to the original temporal structure. For these reasons, the 31-day standardization provides a more reliable and less intrusive alignment strategy for short-term load forecasting.

This expanded and smoothened dataset preserves the natural ebb and flow of consumption across hours, days, and months, enabling the forecasting model to capture seasonal trends with finer precision.

4. Applying Feature Selection

Selecting appropriate predictors is essential for improving model accuracy and avoiding redundant information. Common approaches include constructing a Correlation Matrix, applying Minimum Redundancy Maximum Relevance (MRMR), performing Analysis of Variance (ANOVA), and employing the Chi-Square test. For this analysis, seventeen atmospheric indicators were obtained from the Visual Crossing service, as illustrated in Figure 7. Several reputable weather data providers including the National Weather Service, the Weather Channel, AccuWeather, Weather Underground, and Visual Crossing were examined and compared. Based on a comprehensive assessment of data completeness, spatial coverage, and reliability, Visual Crossing was selected as the primary source of meteorological information, offering the most consistent dataset with 5 min resolution for the case study location.

These variables are: Temperature, the ambient air temperature reported in °C or °F; Apparent Temperature (Feels-Like), a composite measure that accounts for temperature, humidity, and wind chill; Dew Point, the temperature at which air becomes saturated and which indicates moisture content; Relative Humidity, the percentage of water vapor present relative to the maximum possible at that temperature; Precipitation, any form of water liquid or frozen falling from the sky; Precipitation Probability, the likelihood (as a percentage) of precipitation occurring; Snowfall Amount, the depth of snow accumulation; Snow Depth, the total height of snow on the ground; Wind Gust, short bursts of high wind speed exceeding the average; Wind Speed, the sustained wind velocity typically measured in km/h or mph; Wind Direction, the compass bearing (in degrees from true north) from which the wind originates; Sea Level Pressure, atmospheric pressure corrected to sea level; Cloud Cover, the fraction of the sky obscured by clouds expressed as a percentage; Visibility, the maximum distance at which objects or lights can be clearly seen; Solar Radiation, the instantaneous power of sunlight reaching the surface in W/m²; Solar Energy, the cumulative solar irradiance over a specified period in kWh/m²; and UV Index, a standardized scale indicating the strength of ultraviolet radiation.

Because our dataset is both large and subject to irregular fluctuations, it is difficult to discern the interdependencies among variables by simple inspection. Traditional statistical analyses often fall short when faced with the non-linear and multi-factor relationships present in meteorological data. To address this, machine learning algorithms were employed to detect subtle patterns, uncover complex associations, and highlight trends that would otherwise remain hidden.

4.1. Chi-Square Test

The Chi-Square Test [38] is a nonparametric technique used to determine whether two categorical variables are related. In load-forecasting applications, continuous weather measurements are first converted into discrete bins (for example, temperature grouped into “low”, “medium”, and “high”), and energy usage is similarly categorized into consumption ranges. The test then computes a Chi-Square statistic by comparing the observed frequencies in each combination of weather-bin and load-bin with the expected frequencies under the assumption of independence. A larger Chi-Square value indicates a stronger association between a given weather feature and energy consumption.

Procedure:

• Split weather variables (temperature, UV index, etc.) into a set of categories.
• Bin the target: Organize historical load values into energy consumption intervals.
• Compute frequencies: Tally how often each weather-category/load-category pair occurs.
• Calculate statistic: For each feature, sum over all bins using the term:

$${\displaystyle \frac{Observed-Expected}{Expected}}$$

(5)

5.

Rank features: Order variables by their Chi-Square scores higher scores indicate greater predictive value.

When applied to our dataset, the highest scores belong to temperature, feels-like temperature, UV index, sea level pressure, solar energy, and solar radiation (see

Table 3. Results of Chi-Square feature selection.

Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation	Visibility
Score	7162.332	5907.64	4932.46	4775.704	4176.774	4151.809	55.890
Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation
Score	3776.295	2270.17	867.17	676.616	600.80606	284.650

). These results imply that these six weather factors most strongly influence when and how much energy is consumed for heating, cooling, and solar-powered operations. Dew point and humidity produced moderate Chi-Square values, whereas visibility and cloud cover scored relatively low, showing only weak associations with load. All examined variables yielded p-values below the usual significance threshold, confirming that none of these relationships arose by chance.

By selecting top-scoring features such as temperature and solar related variables leaner and more accurate forecasting models can be constructed, while safely omitting the least influential factors.

4.2. Minimum Redundancy Maximum Relevance (MRMR)

The MRMR [39] approach seeks a set of inputs that are both highly informative about the target (energy consumption) and as mutually non-redundant as possible. It does so by ranking features based on mutual information scores, which quantify how much knowing one variable reduces uncertainty about another.

Implementation steps:

• Standardize each numeric feature:

$$StdV={\displaystyle \frac{V-\mu \left(V\right)}{\sigma \left(V\right)}}$$

(6)

where V is the original value, μ(V) its mean, and σ(V) its standard deviation. This puts all features on the same scale, preventing any single unit-scale variable from dominating the MI calculation.

2.

Compute mutual information between each feature and the energy-consumption target.

3.

Rank features by high relevance (large MI) and low redundancy (little overlap with already-selected features).

Table 4 shows the ranked results. In summary:

• Temperature and feels-like temperature emerge as the strongest predictors.
• Dew point, sea level pressure, and humidity have moderate importance.
• Solar energy and visibility register relatively low relevance scores.

Table 4. Results of MRMR feature selection.

Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation	Visibility
Score	0.386	0.381	0.052	0.203	0.073	0.153	0.040
Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation
Score	0.242	0.159	0.139	0.106	0.114	0.085

Table 4. Results of MRMR feature selection.

Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation	Visibility
Score	0.386	0.381	0.052	0.203	0.073	0.153	0.040
Feature	Temperature	Feels-Like	UV Index	Sea Level Pressure	Solar Energy	Solar Radiation
Score	0.242	0.159	0.139	0.106	0.114	0.085

Because building heating and cooling demand is sensitive to temperature-related inputs, those variables drive most of the predictive power. Less-relevant features can be dropped to simplify the model without sacrificing accuracy.

4.3. Correlation Matrix

A correlation matrix (CM) [40] in Figure 8 shows pairwise linear relationships between numeric features and the energy-consumption. Each entry is the Pearson correlation coefficient, ranging from −1 (perfect inverse) through 0 (no linear relationship) to +1 (perfect direct).

The analysis yields the following results:

• Sea level pressure has the strongest negative correlation (−0.34), implying that higher pressures generally coincide with lower energy use (likely due to milder weather).
• Temperature, feels-like temperature, and dew point have strong positive correlations (approximately +0.32, +0.31, and +0.22), reflecting increased cooling demand under warmer or more humid conditions.
• Solar variables (UV index, solar radiation, solar energy) exhibit modest positive correlations (around +0.17).
• Cloud cover and visibility feature near-zero correlations, indicating negligible direct effects on consumption.

This correlation analysis reinforces that temperature and related humidity measures are dominant drivers of energy demand, while factors such as cloud cover and visibility play only minor roles.

The top-ranked features for each selection method are: CM: sea level pressure, temperature, feels-like temperature, and dew point. MRMR: temperature, feels-like temperature, dew point, and sea level pressure. Chi-Square: temperature, feels-like temperature, UV index, and sea level pressure.

Each method serves a different role in the proposed approach and is paired with a different machine learning technique or statistical method.

The proposed framework is not dependent on any specific set of weather variables. All meteorological features used in this study (temperature, relative humidity, sea-level pressure, dew point, UV index, etc.) are publicly available through standard weather services and do not require specialized sensors. The pipeline can operate with smaller or alternative feature sets, depending on what is available in a given region. The feature selection component is included to demonstrate how the model behaves when richer predictor sets exist, but the framework remains fully functional even when only the commonly available variables (e.g., temperature, pressure, humidity, wind, precipitation) are present.

5. Discussion and Comparison of the Results: Overall Consumption Prediction vs. Aggregated Appliance Forecasts

This paper introduced a new cloud-enabled hybrid machine learning–statistical framework for short-term load forecasting in smart buildings. The proposed approach integrates the strengths of both machine learning algorithms and statistical models to maximize forecasting accuracy while simultaneously leveraging the reliability, high processing speed, and cost efficiency of cloud computing. By employing a parallel multi-cloud architecture with a two-out-of-three voting mechanism, the framework achieves even greater speed and reliability compared to conventional single-cloud deployments. This architecture also provides a scalable and fault-tolerant foundation that enables researchers to extend the methodology beyond load forecasting into other critical areas of the power grid, such as fault detection, protection, and predictive maintenance monitoring.

In this study, the next hour of electrical consumption was predicted, corresponding to 12 time steps, with each step representing a 5 min interval. The performance of the proposed method is compared against several well-known forecasting techniques including: Linear Regression (LR), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), SARIMAX, Extreme Gradient Boosting (XGBoost), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Deep Belief Network (DBN). These comparative results are presented in

Table A1. Next 12 predicted (KW) points by different methods: overall consumption.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	1.569	0.911	0.592	1.634	0.853	0.951	0.621	0.938	0.871	0.782	0.650
2	2.104	0.887	0.545	1.610	0.703	0.279	0.602	0.923	0.699	0.775	0.646
3	2.104	1.040	0.516	1.576	0.643	0.513	0.569	0.901	0.637	0.768	0.752
4	1.569	0.941	1.294	1.941	1.963	3.180	0.665	1.112	1.968	1.730	1.840
5	2.104	0.884	0.812	1.658	0.892	0.946	0.637	0.953	0.89	0.883	0.846
6	1.569	0.888	0.580	1.573	0.643	0.513	0.565	0.899	0.643	0.612	0.590
7	1.569	0.920	0.561	1.577	0.631	0.513	0.570	0.902	0.632	0.612	0.590
8	1.569	0.927	0.535	1.624	0.854	0.788	0.613	0.932	0.833	0.708	0.625
9	2.639	1.050	0.535	1.634	0.853	0.951	0.621	0.938	0.865	0.795	0.664
10	2.104	0.958	1.247	2.004	1.708	1.031	0.642	1.143	1.71	1.714	1.732
11	2.639	0.902	0.789	1.689	0.988	1.019	0.655	0.972	0.988	0.989	0.989
12	1.569	0.911	0.575	1.596	0.738	0.794	0.589	0.914	0.737	0.701	0.699

,

Table A2. Next 12 predicted (KW) points by different methods: FAU (forced air unit) downstairs.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.017513	0.008902	0.007134	0.012425	0.006982	0.009112	0.007698	0.008235	0.007091	0.006785	0.00601
2	0.018425	0.009503	0.00721	0.012833	0.006991	0.009621	0.007895	0.008501	0.007254	0.006945	0.006043
3	0.017901	0.007625	0.007625	0.012332	0.006951	0.009432	0.00921	0.007193	0.008395	0.006851	0.005963
4	1.202331	0.731299	0.557128	0.983572	0.645221	0.71294	0.58932	0.638521	0.552118	0.501237	0.477047
5	0.391254	0.251332	0.173555	0.288413	0.210892	0.239144	0.199813	0.220335	0.176624	0.161225	0.145373
6	0.018311	0.009108	0.007324	0.012641	0.006914	0.009405	0.007809	0.00829	0.007114	0.006831	0.006037
7	0.018812	0.00961	0.007501	0.013001	0.007225	0.009832	0.008014	0.008502	0.007334	0.007009	0.006147
8	0.018501	0.009215	0.007254	0.012721	0.007001	0.009611	0.007921	0.008411	0.007209	0.006982	0.00611
9	0.0189	0.0097	0.007605	0.01322	0.007332	0.009955	0.008225	0.00872	0.007445	0.00712	0.00617
10	1.210422	0.740291	0.56321	0.990512	0.653222	0.721008	0.59512	0.642892	0.557231	0.503188	0.477707
11	0.39012	0.250992	0.172801	0.287134	0.209921	0.238551	0.198923	0.219832	0.175832	0.160534	0.145263
12	0.01825	0.00902	0.007111	0.01252	0.006952	0.009255	0.007715	0.008175	0.007003	0.006822	0.006033

,

Table A3. Next 12 predicted (KW) points by different methods: air conditioning downstairs.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.002301	0.001205	0.000924	0.00161	0.000851	0.001132	0.000967	0.001031	0.000912	0.000843	0.00078
2	0.002298	0.001211	0.000919	0.001599	0.000845	0.001127	0.000963	0.001028	0.000908	0.000839	0.00078
3	0.002115	0.001099	0.000842	0.001503	0.000799	0.001046	0.000892	0.00095	0.000838	0.000772	0.0007
4	0.002069	0.001067	0.000826	0.001469	0.00078	0.001021	0.00087	0.000928	0.00082	0.000755	0.00069
5	0.001986	0.001025	0.000793	0.001409	0.000748	0.000979	0.000834	0.00089	0.000787	0.000724	0.00066
6	0.002083	0.001074	0.00083	0.001453	0.000772	0.001009	0.00086	0.000919	0.000815	0.000747	0.000693
7	0.002352	0.001215	0.000938	0.001644	0.000869	0.001154	0.000985	0.001051	0.000933	0.000865	0.00081
8	0.001858	0.000961	0.000742	0.001298	0.000688	0.000906	0.000774	0.000827	0.000737	0.000676	0.000643
9	0.002067	0.001072	0.00083	0.00145	0.00077	0.001006	0.000858	0.000917	0.000813	0.000746	0.00071
10	0.002157	0.00112	0.000865	0.001515	0.000805	0.001051	0.000897	0.000958	0.000851	0.000781	0.00074
11	0.002276	0.001183	0.000914	0.001619	0.000861	0.001124	0.00096	0.001028	0.000913	0.000845	0.000773
12	0.002296	0.001196	0.000922	0.001635	0.000871	0.001137	0.00097	0.001039	0.000921	0.000852	0.000783

,

Table A4. Next 12 predicted (KW) points by different methods: AC upstairs.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	−0.00108	−0.00052	−0.00041	−0.0008	−0.00033	−0.00049	−0.00045	−0.00047	−0.00034	−0.0003	−0.00036
2	−0.00106	−0.00051	−0.0004	−0.00078	−0.00033	−0.00049	−0.00044	−0.00047	−0.00033	−0.00029	−0.00035
3	−0.00113	−0.00054	−0.00043	−0.00082	−0.00034	−0.00051	−0.00046	−0.00049	−0.00035	−0.00031	−0.00038
4	−0.00137	−0.00066	−0.00052	−0.00099	−0.00042	−0.00062	−0.00056	−0.00059	−0.00042	−0.00037	−0.00046
5	−0.00125	−0.0006	−0.00047	−0.0009	−0.00038	−0.00056	−0.00051	−0.00054	−0.00038	−0.00034	−0.00042
6	−0.00123	−0.00059	−0.00046	−0.00089	−0.00037	−0.00055	−0.0005	−0.00053	−0.00038	−0.00033	−0.00041
7	−0.00097	−0.00046	−0.00036	−0.0007	−0.00029	−0.00044	−0.00039	−0.00042	−0.0003	−0.00026	−0.00032
8	−0.00142	−0.00068	−0.00054	−0.00104	−0.00043	−0.00064	−0.00058	−0.00062	−0.00044	−0.00038	−0.00048
9	−0.00161	−0.00077	−0.00061	−0.00117	−0.00049	−0.00072	−0.00065	−0.00069	−0.00049	−0.00043	−0.00054
10	−0.00164	−0.00079	−0.00062	−0.00119	−0.0005	−0.00074	−0.00067	−0.00071	−0.0005	−0.00044	−0.00055
11	−0.00134	−0.00064	−0.00051	−0.00098	−0.00041	−0.00061	−0.00055	−0.00059	−0.00042	−0.00036	−0.00045
12	−0.00155	−0.00074	−0.00059	−0.00113	−0.00047	−0.0007	−0.00063	−0.00067	−0.00048	−0.00042	−0.00052

,

Table A5. Next 12 predicted (KW) points by different methods: refrigerator.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSXLR	Actual Data
1	0.018945	0.009118	0.007204	0.013375	0.007001	0.009402	0.008015	0.008465	0.007293	0.006981	0.006727
2	0.018832	0.009065	0.007166	0.013288	0.006942	0.009351	0.007984	0.008432	0.007259	0.006951	0.006703
3	0.01884	0.009071	0.00717	0.013296	0.006946	0.009356	0.007988	0.008437	0.007263	0.006955	0.006703
4	0.017858	0.008604	0.00681	0.01264	0.006601	0.008892	0.0076	0.008042	0.006897	0.006581	0.006343
5	0.018201	0.008771	0.00694	0.012894	0.00674	0.009086	0.007769	0.00823	0.00706	0.006732	0.00648
6	0.01826	0.0088	0.006962	0.012933	0.006761	0.009115	0.007794	0.008255	0.007081	0.006751	0.0065
7	0.018273	0.008807	0.006968	0.012944	0.006768	0.009122	0.0078	0.008261	0.007087	0.006756	0.0065
8	0.018284	0.008812	0.006971	0.01295	0.006771	0.009126	0.007804	0.008265	0.00709	0.006759	0.006503
9	0.018323	0.008832	0.006986	0.012978	0.006788	0.009145	0.00782	0.00828	0.007105	0.006772	0.006517
10	0.017671	0.008507	0.006718	0.012506	0.00655	0.008794	0.007527	0.007974	0.006833	0.006528	0.006193
11	0.018766	0.00902	0.007139	0.013233	0.006915	0.009312	0.007955	0.008407	0.007238	0.006931	0.006593
12	0.0189	0.009091	0.007201	0.01335	0.006986	0.009381	0.00801	0.00846	0.00729	0.006981	0.00664

,

Table A6. Next 12 predicted (KW) points by different methods: laundry/dryer.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.010825	0.005273	0.004242	0.007917	0.004008	0.005536	0.004821	0.005129	0.004181	0.003956	0.003603
2	0.01091	0.005316	0.004277	0.007988	0.004045	0.005588	0.004866	0.005175	0.004217	0.00399	0.003637
3	0.010789	0.005256	0.004227	0.007878	0.003993	0.005522	0.004804	0.005113	0.00416	0.003935	0.00358
4	0.010833	0.005278	0.004244	0.007913	0.004012	0.005539	0.004822	0.005133	0.004176	0.003951	0.003607
5	0.010537	0.005131	0.004127	0.00769	0.003898	0.005382	0.004686	0.004983	0.004058	0.003836	0.00351
6	0.010678	0.005202	0.004182	0.007824	0.003964	0.005464	0.004762	0.005063	0.004118	0.003893	0.003563
7	0.010579	0.005152	0.004142	0.007738	0.003921	0.005414	0.004718	0.005018	0.004079	0.003856	0.003527
8	0.010584	0.005154	0.004145	0.007744	0.003924	0.005418	0.004721	0.005021	0.004082	0.003858	0.003527
9	0.010606	0.005166	0.004154	0.00776	0.003934	0.00543	0.004731	0.005032	0.004091	0.003867	0.00352
10	0.010582	0.005154	0.004144	0.007742	0.003923	0.005417	0.00472	0.00502	0.004081	0.003857	0.003527
11	0.010919	0.00532	0.004281	0.008003	0.00405	0.005593	0.004872	0.005183	0.004224	0.003996	0.003637
12	0.010732	0.005227	0.004207	0.007853	0.003972	0.00549	0.004787	0.00509	0.004142	0.003919	0.003567

,

Table A7. Next 12 predicted (KW) points by different methods: dishwasher/garbage disposal.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.002256	0.001121	0.000874	0.001661	0.000752	0.001051	0.000942	0.001003	0.000789	0.000705	0.00075
2	0.002452	0.001222	0.000953	0.001799	0.000821	0.001152	0.00103	0.001098	0.000863	0.000773	0.000817
3	0.002224	0.001103	0.000859	0.001636	0.00074	0.001033	0.000926	0.000985	0.000775	0.000693	0.00074
4	0.004831	0.002374	0.001857	0.003555	0.001598	0.002243	0.002026	0.002155	0.001692	0.001515	0.001643
5	0.003074	0.001514	0.001186	0.002263	0.001022	0.001435	0.001296	0.001379	0.001082	0.000968	0.001047
6	0.002365	0.001174	0.000913	0.001719	0.000777	0.001093	0.00098	0.001049	0.000823	0.000739	0.000787
7	0.002599	0.001283	0.000997	0.001892	0.000857	0.0012	0.001077	0.001151	0.000902	0.000809	0.000863
8	0.002129	0.00105	0.000814	0.001552	0.000704	0.000985	0.000882	0.000944	0.00074	0.000664	0.000707
9	0.002097	0.001034	0.000802	0.001528	0.000693	0.00097	0.000869	0.00093	0.000729	0.000654	0.000697
10	0.004711	0.002321	0.001815	0.003466	0.001556	0.002184	0.001974	0.0021	0.00165	0.001477	0.00161
11	0.002857	0.001414	0.001106	0.002102	0.000945	0.001326	0.00119	0.001272	0.001	0.000897	0.000977
12	0.002264	0.00112	0.000874	0.001662	0.000752	0.00105	0.000942	0.001004	0.00079	0.000705	0.00075

,

Table A8. Next 12 predicted (KW) points by different methods: microwave.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.003218	0.001858	0.001458	0.002322	0.001188	0.001566	0.001501	0.001598	0.001382	0.001318	0.00108
2	0.003312	0.001927	0.00151	0.002387	0.001239	0.001645	0.001555	0.001678	0.001432	0.001362	0.0011266
3	0.003099	0.001774	0.001387	0.002224	0.001138	0.001512	0.001429	0.001535	0.001316	0.001252	0.0010433
4	0.003178	0.001847	0.001439	0.002299	0.001191	0.001568	0.001481	0.001589	0.001364	0.00131	0.0010733
5	0.003275	0.001898	0.001487	0.002364	0.001221	0.00161	0.001532	0.001654	0.00141	0.001343	0.00111
6	0.003214	0.001864	0.001459	0.002327	0.001206	0.00159	0.001503	0.001624	0.001395	0.001339	0.0010966
7	0.003329	0.00192	0.001507	0.002402	0.001228	0.001631	0.001552	0.001665	0.001428	0.001363	0.0011166
8	0.003112	0.001814	0.001417	0.002265	0.001169	0.001546	0.00146	0.001577	0.001342	0.001298	0.0010733
9	0.0032	0.001843	0.001432	0.002295	0.001184	0.001561	0.001474	0.001594	0.001369	0.001313	0.0010766
10	0.003252	0.001892	0.001481	0.002362	0.001225	0.001616	0.001536	0.001637	0.001403	0.001347	0.0011133
11	0.003284	0.00191	0.001497	0.002389	0.001228	0.001631	0.00154	0.001653	0.00142	0.001363	0.0011166
12	0.003286	0.001909	0.001499	0.002387	0.001232	0.001634	0.001543	0.001654	0.001421	0.001354	0.00111

,

Table A9. Next 12 predicted (KW) points by different methods: FAU upstairs.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.023895	0.01113	0.009172	0.017882	0.008612	0.012544	0.011276	0.011869	0.009442	0.008772	0.00794
2	0.02366	0.011016	0.00908	0.017737	0.008532	0.012452	0.011191	0.01178	0.009363	0.008698	0.007887
3	0.02375	0.011062	0.009115	0.017801	0.008568	0.012493	0.011227	0.011821	0.009392	0.008728	0.007917
4	1.84365	0.869225	0.694118	1.385427	0.657311	0.951882	0.864113	0.910225	0.702447	0.635822	0.62757
5	0.320535	0.155773	0.124193	0.244153	0.115867	0.168152	0.152559	0.161037	0.12491	0.112967	0.109373
6	0.02404	0.011191	0.009218	0.017975	0.008665	0.012605	0.011332	0.011929	0.009491	0.008819	0.008013
7	0.02371	0.011043	0.009107	0.017768	0.008554	0.012476	0.011212	0.011802	0.009376	0.008713	0.007903
8	0.02399	0.011177	0.009207	0.01795	0.008655	0.012595	0.011323	0.01192	0.009481	0.00881	0.008007
9	0.02387	0.011118	0.00916	0.017868	0.008608	0.012544	0.011276	0.011871	0.009437	0.008771	0.007967
10	1.84782	0.871025	0.695452	1.388362	0.659014	0.954193	0.866174	0.912356	0.704123	0.63731	0.629273
11	0.317834	0.154462	0.123164	0.242273	0.114952	0.166883	0.151423	0.159822	0.123991	0.112142	0.10793
12	0.023565	0.010981	0.00905	0.017696	0.008513	0.012426	0.01117	0.011758	0.009338	0.008672	0.00785

,

Table A10. Next 12 predicted (KW) points by different methods: kitchen GFI (ground fault circuit interrupter)/toaster.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	0.000033	0.000018	0.000014	0.000027	0.000012	0.000019	0.000017	0.000018	0.000014	0.000013	0.0000133
2	0.00006	0.00003	0.000024	0.000048	0.000021	0.000033	0.000029	0.000031	0.000024	0.000022	0.00002
3	0.000108	0.000052	0.000041	0.000086	0.000038	0.000058	0.000051	0.000055	0.000043	0.000039	0.0000367
4	0.000176	0.000084	0.000067	0.00014	0.000061	0.000095	0.000084	0.00009	0.000071	0.000065	0.00006
5	0.000168	0.00008	0.000064	0.000134	0.000059	0.00009	0.000081	0.000087	0.000069	0.000063	0.0000567
6	0.00017	0.000081	0.000065	0.000136	0.000059	0.000091	0.000082	0.000088	0.00007	0.000064	0.0000567
7	0.00013	0.000062	0.00005	0.000104	0.000045	0.000069	0.000063	0.000067	0.000053	0.000048	0.0000433
8	0.00033	0.000155	0.000124	0.000264	0.000118	0.000176	0.000159	0.000168	0.000133	0.00012	0.00011
9	0.000249	0.00012	0.000095	0.0002	0.000089	0.000133	0.00012	0.000127	0.000101	0.000092	0.0000833
10	0.000289	0.000139	0.000111	0.000232	0.000103	0.000154	0.000139	0.000147	0.000116	0.000105	0.0000967
11	0.000219	0.000106	0.000084	0.000176	0.000079	0.000118	0.000106	0.000112	0.000089	0.000081	0.0000733
12	0.00009	0.000044	0.000035	0.000072	0.000032	0.00005	0.000045	0.000047	0.000037	0.000034	0.00003

and

Table A11. Next 12 predicted (KW) points by different methods: general (miscellaneous) household load.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	1.822104	0.91924	0.676118	1.392781	0.210125	0.534925	0.61302	0.57248	0.225194	0.196833	0.624233
2	1.807326	0.910481	0.669839	1.38035	0.208012	0.529689	0.607623	0.56747	0.223217	0.195084	0.619827
3	2.1154	1.06139	0.780626	1.61483	0.243584	0.620552	0.71228	0.665982	0.261883	0.229029	0.72587
4	2.10937	1.05824	0.778301	1.610066	0.242905	0.618736	0.710184	0.664031	0.260815	0.228099	0.723123
5	1.69062	0.848426	0.624025	1.29016	0.194717	0.495586	0.568477	0.530534	0.208309	0.182078	0.579567
6	1.64558	0.825861	0.60728	1.255844	0.189549	0.482486	0.553916	0.51662	0.203005	0.177666	0.56406
7	1.64375	0.82494	0.606606	1.254294	0.189313	0.481886	0.553226	0.515967	0.202685	0.177383	0.563447
8	1.74768	0.877021	0.644677	1.33329	0.201133	0.511807	0.587489	0.547783	0.21495	0.18815	0.59914
9	1.86149	0.933968	0.685921	1.42944	0.215455	0.547872	0.628909	0.586144	0.231409	0.202521	0.63872
10	1.78459	0.894606	0.657066	1.3691	0.206324	0.524664	0.602642	0.561501	0.220727	0.193142	0.612323
11	2.11404	1.06072	0.780099	1.61376	0.243448	0.620224	0.711905	0.665644	0.261753	0.228913	0.723473
12	1.96819	0.988808	0.726622	1.50408	0.226812	0.578106	0.663648	0.619833	0.243294	0.213253	0.673247

(in the Appendix A), highlighting the improvements achieved through the dual-stage trend–remainder modeling process in combination with the robust cloud-based deployment.

Table A1 and Figure 9: Overall consumption, Table A2 and Figure 10: FAU (Forced Air Unit) Downstairs consumption, Table A3 and Figure 11: AC (Air Conditioning) Downstairs consumption, Table A4 and Figure 12: Refrigerator consumption, Table A5 and Figure 13: AC Upstairs consumption, Table A6 and Figure 14: Laundry/Dryer consumption, Table A7 and Figure 15: Dishwasher/Garbage Disposal consumption, Table A8 and Figure 16: Microwave consumption, Table A9 and Figure 17: FAU Upstairs consumption, Table A10 and Figure 18: Kitchen GFI (Ground Fault Circuit Interrupter)/Toaster consumption, Table A11 and Figure 19: General (miscellaneous) load consumption and

Table A12. Next 12 predicted (KW) points by different methods: overall consumption—summation of Table A4, Table A5, Table A6, Table A7, Table A8, Table A9, Table A10, Table A11 and Table A12.

Method Point	DBN	LSTM	SARIMAX	LR	RF	DT	SVM	GRU	XGBOOST	SSRLR	Actual Data
1	4.402809	0.996231	1.087194	4.957531	0.753281	0.924844	1.001538	0.829909	1.033635	0.740985	0.650
2	4.527383	1.022295	1.101331	5.090235	0.760627	0.940943	1.015361	0.844171	1.040605	0.747044	0.646
3	4.838178	1.008571	1.225377	5.306084	0.844195	1.01274	1.083204	0.918764	1.162196	0.831497	0.752
4	9.84792	1.645598	1.992939	9.40809	2.022073	2.111469	2.054901	1.659253	2.045687	2.183267	1.840
5	5.580026	1.113722	1.2569	5.837724	0.982709	1.214348	1.236753	1.003334	1.209144	1.024185	0.846
6	4.546716	0.976813	1.05935	5.12311	0.727287	0.904705	0.968167	0.813176	0.986795	0.709672	0.590
7	4.611737	1.04493	1.069944	5.27021	0.740032	0.926468	1.005123	0.829526	0.997419	0.719465	0.590
8	4.975247	0.991344	1.151328	5.511958	0.778107	0.95712	1.009119	0.865246	1.056523	0.755695	0.625
9	4.854031	1.004369	1.159956	5.392624	0.789248	0.961816	1.023475	0.870084	1.078246	0.773757	0.664
10	9.811275	1.671429	1.906255	9.498042	1.959133	2.057394	1.998976	1.598688	1.939943	2.111271	1.732
11	6.211354	1.246643	1.465467	6.482264	1.119608	1.35283	1.388854	1.136581	1.403961	1.159455	0.989
12	4.606389	1.00431	1.149319	5.153972	0.791093	0.961008	1.034905	0.868796	1.091916	0.783885	0.699

and Figure 20: Sum of results of Table A2, Table A3, Table A4, Table A5, Table A6, Table A7, Table A8, Table A9, Table A10 and Table A11.

It is important to note that occasional negative consumption values may appear in the dataset. These values do not indicate that appliances such as air conditioners are generating energy; rather, they are a result of the building’s integration with rooftop solar panels. During periods of high solar production particularly when household energy demand is low excess electricity is exported back to the grid.

Five well-established evaluation metrics (

Table 5. Evaluation Metrics and Formulas.

Metric	Equation
MSE [41]	MSE = ${\displaystyle \frac{1}{N}}\sum _{i=1}^N{(A_i-F_i)}^2$
MAE [42]	MAE = ${\displaystyle \frac{1}{N}}\sum _{i=1}^N\left|F_i-A_i\right|$
MAPE [43]	MAPE = ${\displaystyle \frac{1}{N}}\sum _{i=1}^N\left|{\displaystyle \frac{A_i-F_i}{A_i}}\right|\times 100\%$
RMSE [44]	RMSE = $\sqrt{{\displaystyle \frac{1}{N}}\times \sum _{i=1}^N{(F_i-A_i)}^2}$
RMSLE [45]	RMSLE = $\sqrt{{\displaystyle \frac{1}{N}}\times \sum _{i=1}^N{({\mathrm{l}\mathrm{o}\mathrm{g}(F}_i+1)}^2-{log(A_i+1))}^2}$

) were employed to compare the results. Since all models were evaluated on the same single-building dataset, normalization measures (e.g., cvRMSE, NMBE) were not applied, as they become relevant only in multi-building comparisons, in accordance with ASHRAE Guideline 14 and IPMVP.

Although the model is trained and evaluated at a 5 min resolution, the output for each horizon is a 12-point vector covering the next hour. For deployment, building EMS/DR (EMS = Energy Management System, DR = Demand Response) controllers can map these to their operating cadence by forming block averages (10/15/30/60 min) or energy-integrated setpoints. This hierarchical use of the same forecasts attenuates short-term noise and matches actuator latencies, while retaining sub-hour detail for rapid interventions (e.g., staggered EV charging). Integrating the proposed forecasting framework with demand response (DR) controllers or optimization models could therefore enhance load flexibility, grid stability, and energy-cost savings an aspect identified as a promising direction for future work.

Aggregate vs. Sum of Appliances:

On the aggregate meter, SSRLR is the best model on every metric (RMSLE = 0.04311, RMSE = 0.07807, MAE = 0.05850). The next best RMSLE is RF = 0.06610, so SSRLR improves AF, RMSLE by ~34.8%.

For the sum of appliance forecasts, the best model is RF with RMSLE = 0.07128 (MAE = 0.13657, RMSE = 0.14152) and second is the SSRLR with RMSLE = 0.08009. Comparing the two strategies best vs. best, AF: SSRLR (0.04311) beats sum of appliances RF (0.07128) by ~39.5% in RMSLE (the summer series is ~65% worse). This behavior is expected: when many appliances operate intermittently (e.g., microwave, dishwasher, small plugs), their individual errors can accumulate when the loads are summed, even if each model performs well.

In short, direct aggregate forecasting uses stable daily/weekly patterns in the total load, while the AL approach must first learn many on/off behaviors and then add them. Without hierarchical reconciliation (to force the sum to match the meter) or special handling of zeros/events, the AL sum can penalize small timing errors and stack relative errors.

Results of Per Appliance Forecasting:

Averaged across all appliances, SSRLR ranks first (avg RMSLE ≈ 0.042; RMSE/MAE show the same trend). RF and XGBoost are the strongest baselines. SSRLR is the best method for every single appliance; examples:

• FAU Downstairs: RMSLE 0.01783, RMSE 0.02536
• AC Downstairs: RMSLE 0.04838, RMSE 0.04962
• Refrigerator: RMSLE 0.04195, RMSE 0.04103
• AC Upstairs: RMSLE 0.04732, RMSE 0.04878
• Laundry/Dryer: RMSLE 0.04137, RMSE 0.04238
• Dishwasher/Garbage: RMSLE 0.04063, RMSE 0.04165
• Microwave: RMSLE 0.04116, RMSE 0.04207
• FAU Upstairs: RMSLE 0.04192, RMSE 0.07004
• Kitchen GFI/Toaster: RMSLE 0.03949, RMSE 0.04051
• General/Misc: SSRLR is best but errors are higher (RMSLE 0.06854, RMSE 0.10892) because this circuit mixes many small, irregular uses.

By the appliance type:

• Spiky/episodic loads (Microwave, Dishwasher/Garbage, Laundry) are handled well by SSRLR, which efficiently manages short bursts.
• HVAC loads (AC and FAU) also favor SSRLR, showing that it captures both seasonality and start/stop behavior.
• General/Misc is the hardest because it blends many small and random loads: SSRLR shows the best accuracy.

Model Ranking and Error Profiles:

A consistent ranking appears across aggregate, appliances, and the sum:

• SSRLR is best overall (aggregate and per appliance averages), strong on both relative (RMSLE) and absolute (RMSE/MAE) errors.
• RF/XGBoost are reliable runner ups; they capture nonlinear effects and are stable across devices.
• SARIMAX is a good seasonal baseline, but weaker on sharp spikes.
• LSTM and GRU do not dominate here, likely due to intermittency, short forecast horizons, and limited effective data per appliance.
• LR/DT/DBN/SVM perform poorly for short-term load forecasting in this setting.

This order holds for the aggregate and appliance averages. In the sum, RF wins, but it still trails the aggregate SSRLR by a wide margin.

Key Takeaways:

• Direct aggregate forecasting (SSRLR) is much more accurate than the sum of appliance forecasts on this dataset (~39.5% lower RMSLE, best vs. best).
• SSRLR is the most consistent model at both levels; RF/XGBoost are strong baselines.

It may look surprising that SSRLR wins for every appliance, yet RF wins after summing the results. The reason for that is the shift of the error pattern when summing consumption of many appliance forecasts: small positively correlated residuals from several SSRLR appliance models can add up for the same time points (for example, around evening peaks). RF tends to produce smoother appliance predictions with less shared timing error across devices, so when added some errors cancel each other. In addition, hierarchical reconciliation was not applied; therefore, perfect per-device calibration does not guarantee that the aggregated total matches the main meter. Because RMSLE on the summed series is sensitive to synchronized timing errors, RF ends up with a lower summed RMSLE (0.07128) than SSRLR (0.08009). Adding reconciliation and aggregate aware training would likely reduce and may remove this gap.

6. Conclusions

This paper introduced a cloud-enabled hybrid ML–statistical framework for 5 min residential load forecasting that blends Seasonal ARIMAX, Random Forest, and LSTM within a residual-correction architecture (SSRLR) and deploys redundant instances across multiple clouds with a two-out-of-three voting scheme. Using sub-metered data from a residential home in Summerlin, Las Vegas (2022), aggregate forecasting was evaluated against appliance-level forecasting. The evidence is clear: on the aggregate series, SSRLR is best across all metrics (RMSLE = 0.04311, RMSE = 0.07807, MAE = 0.05850), outperforming the next-best model (RF, RMSLE = 0.06610) by ~34.8%. For the appliance-level results, the RF model performs best (RMSLE = 0.07128), followed by the SSRLR model (RMSLE = 0.08009). However, the aggregate SSRLR model still outperforms the appliance-level results by up to approximately 39.5% in terms of RMSLE. This gap aligns with expectations for many intermittent end-uses: per appliance models can produce positively correlated residuals that accumulate when added, while the aggregate model exploits stable daily/weekly total load patterns. Importantly, RMSLE is the most informative headline metric in this setting because many appliances operate near zero and fairly scores small absolute errors; RMSE/MAE confirm the same ranking in absolute terms.

Practically, if the goal is maximum accuracy with minimal complexity, a single SSRLR model for the aggregate prediction is the preferred choice. The proposed approach supports real-time, fault-tolerant forecasting for smart-city energy management. Its cloud redundancy mechanism ensures operational continuity during single-cloud failures, making it suitable for deployment in distributed residential and micro-grid environments. The framework’s modular design allows utilities to balance accuracy, latency, and cost by tuning cloud resources and model complexity.

Future work will extend the framework to multi-household, high-load buildings and cross-climate datasets, explore transfer learning for unseen dwellings, and integrate edge-cloud coordination for improved energy-efficiency and privacy. Further research will also focus on adaptive weighting and strengthen interpretability for decision-support systems.

Overall, this research demonstrates that hybrid multi-cloud forecasting can provide fine-grained, resilient, and scalable solutions for next-generation smart-residential infrastructures.