Multi-Horizon Machine Learning Model for Home Energy Consumption Prediction Considering Appliances Usage

Proposed prediction models

a.

The architectures of the proposed models are illustrated in Figures 1–5. It consists of five schemas with respect to the types of power prediction and input feature groups, as follows:

1. Scheme-1: The individual appliance power prediction.

2. Scheme-2: Total power prediction with Feature Group-A (water heater [WH] power, washing machine [WM] power, rice cooker [RC] power, microwave oven [MO] power, total power).

3. Scheme-3: Total power prediction with Feature Group-B (refrigerator [RFG] power, freezer [FZ] power, water pump [WP] power, lighting [LG] power, total power).

4. Scheme-4: Total power prediction with Feature Group-C (water heater power, washing machine power, RC power, MO power, RFZ power, FZ power, WP power, lighting power, total power).

5. Scheme-5: Total power prediction with Feature Group-D (total power).

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Scheme-1 is used to predict the individual appliance power at a future time based on its previous time power consumption. In this work, eight home appliances are considered, namely water heater, washing machine, rice cooker, MO, refrigerator, freezer, WP, and lighting. It is worth noting that with the recent development of IoT technology, individual appliance power consumption monitoring can be implemented easily, as developed in [24, 25].

Scheme-2 to Scheme-5 is used to predict the total power at a future time based on the different input features at previous times. Scheme-2 uses Feature Group-A, consisting of the reschedulable loads WH power, WM power, RC power, MO power and the total power, as the input features. The reschedulable load is the load that the user can control its operating time. Thus, Scheme-2 can be used to examine the impact of occupant behavior when using appliances on the prediction of total power consumption.

Scheme-3 uses Feature Group-B, consisting of the non-reschedulable loads RFG power, FZ power, WP power, LG power and the total power, as the input features. The non-reschedulable load is the load that the user cannot control its operating time. Thus, Scheme-3 can be used to examine the impact of the non-reschedulable appliances on the prediction of total power consumption.

Scheme-4 uses Feature Group-C, consisting of the reschedulable loads, non-reschedulable loads, and the total power, as the input features. Thus, Scheme-4 can be used to examine the impact of all individual appliances on the prediction of total power consumption.

Scheme-5 uses Feature Group-D, consisting solely of total power, as the input feature. Thus, Scheme-5 can be used to examine the performance of total power consumption prediction when using a single feature.

In each scheme, four different time horizons are examined, that is, (a) a 1-min time horizon with a 1-min time interval and a 5-min time step; (b) a 15-min time horizon with a 1-min time interval and a 15-min time step; (c) a 1-hr time horizon with a 5-min time interval and a 6-hr time step; (d) a 1-day time horizon with a 5-min time interval and a 2-day time step. These time-horizon variations fall under VSTLF and STLF. This selection is suitable for analyzing the real-time and hourly or daily operation of home appliances.

As shown in Figures 1–5, six prediction methods are evaluated in each scheme. The prediction techniques are divided into two categories: tree-based machine learning algorithms (RF, LGBM, and XGB) and deep learning-based algorithms (RNN, LSTM, and GRU). The algorithms are described briefly in the following section. The RF, LGBM, and XGB algorithms are usually used for predicting tabular, non-sequential data. Meanwhile, RNNs, LSTMs, and GRUs are commonly used for predicting sequential and time-series data. By evaluating both approaches, the suitability of the algorithm for predicting the short-term power consumption of individual appliances and the total power can be comprehensively analyzed.

Tree-based machine learning algorithms

b.

The tree-based machine learning algorithms used in this work are RF, LGBM, and XGB. They are popular and influential tree-based algorithms used for prediction. RF is a tree-based ensemble machine learning technique developed by [26]. It uses trees as the base learners [27] denoted by h_j (X,Θ_j), where X represents the real-value input of predictor, Θ_j is a collection of random variables. The algorithm is as follows:

1. Prepare training set TS = {(x₁, y₁),(x₂, y₂),...,(x_N, y_N)}, where x_i = (x_i.1,..., x_i,p)^T is the p predictors, and y_i is the response;

2. For j = 1 to J (J is the number of regressors), repeat Step-3 to Step-4;

3. Pick a bootstrap sample TS_j of size N from the training set TS;

4. Fit a tree using the bootstrapped sample TS:

   • (a) Start with all observations on a single node

   • (b) Randomly select the m predictors from the available p predictors

   • (c) Find the best split from the m predictors

   • (d) Split the node into two child nodes using the best split

   • (e) Repeat (b) – (d) until the stopping criterion is met: The prediction of the new point x is given by

The LGBM is a novel gradient boosting machine (GBM) algorithm [28] developed by [29]. It proposes the gradient-based one-side sampling (GOSS) and exclusive feature bundling (EFB) to reduce the computation time. The GOSS reduces the number of samples by focusing on samples with larger gradients and performing random sampling on those with smaller gradients. At first, it sorts the data samples according to the absolute value of the gradients, and selects the highest a × 100% samples. The algorithm then randomly selects b × 100% samples from the rest of data. Finally, the information gain is calculated by amplifying the sampled data with small gradients by a factor $\frac{1-a}{b}$ . Meanwhile, the EFB is used to bundle exclusive features into a single feature, thus reducing the histogram complexity.

XGB [30] is the most popular implementation of the GBM [31, 32]. The algorithm is briefly described in the following. Suppose the dataset DS with m features and n number of samples is expressed as DS = {(x_i, y_i) : i = 1,...,n, x_i ∈ R^m, y_i ∈ R}. The predicted output ŷ_i of an ensemble tree model is given by

The second-order Taylor polynomial can be used to replace the function f_t. Therefore, the objective function becomes

The constant terms in Eq. (6) can be removed to simplify the objective function as follows:

By expanding Ω, the objection function in Eq. (9) can be rewritten as

Deep learning-based prediction techniques

c.

The deep learning-based prediction techniques used in this work include RNNs, LSTMs, and GRUs. They are popular prediction algorithms suitable for time-series data. A RNN [33] is a type of ANN in which the output of a neuron is looped back to the same neuron in the next time step. Thus, it can memorize the previous input data. A RNN cell is illustrated in Figure 6, where the hidden layer state at time t (h_t) is expressed as

Figure 6:

The main drawbacks of RNN are the vanishing and exploding gradient problems, in which the gradient becomes too small (vanishing) or too large (exploding). To overcome these problems, the LSTM was proposed by [34]. The LSTM cell is depicted in Figure 7. It consists of an input gate, a forget gate, and an output gate. The input gate is expressed using the following formula:

Figure 7:

The GRU proposed by [35] is a simplified version of LSTM that improves the performance with less training time. The GRU cell is depicted in Figure 8. The GRU has only two gates, namely the update and reset gates. The update gate is expressed using the following formula:

Figure 8:

Algorithm implementation and evaluation metric

d.

The prediction algorithms are implemented in Python, with machine learning libraries and frame-works employed. The RF is implemented using the scikit-learn library [36]. The LGBM is implemented using the LGBM library [37]. The XGB is implemented using the XGB library [38]. The RNN, LSTM, and GRU are implemented using the Tensorflow framework [39].

The performance of the prediction algorithms are evaluated using two metrics: Normalized mean absolute error (NMAE) and normalized root mean square error (NRMSE). The NMAE and NRMSE are expressed using the following formulas:

Dataset and machine learning parameters

a.

The dataset was collected from a home using an IoT-based SCADA monitoring system, where individual appliances are measured by smart meters connected to the HMI SCADA wirelessly [25]. The HMI/SCADA system was equipped with local data storage and can be accessed remotely via the internet. It is worth noting that this configuration provided a simple, low-cost home appliance power-monitoring system.

The data were collected over a month, from August 1, 2025 to August 31, 2025, with a 1-min time interval. The dataset consisted of timestamps and power consumption for home appliances (e.g., WH, WM, RC, MO, RFG, FZ, WP, LG), as well as the total power consumption. The dataset is saved in the comma-separated value format. The dataset for 1 month contains 44,637 data points.

To ensure accurate and stable predictions, the dataset is cleaned and preprocessed, including handling missing values, outliers, and normalizing the data to a consistent scale. Then, the data are split into 0.75 for training and 0.25 for testing.

The machine learning parameters used in this work are listed in Table 1 for tree-based algorithms and in Table 2 for deep learning algorithms.

Evaluation of individual appliance power prediction

b.

Figures 9–12 depict the NMAE of individual appliance power predictions for 1-min, 15-min, 1-hr, and 1-day time horizons. In each figure, the NMAEs for six methods are compared across eight appliances. To facilitate analysis, the comparison results are summarized in Table 3. Table 3 shows that the lowest NMAE of 0.09% is achieved by the MO using an RNN at a 15-min horizon for reschedulable load, and the lowest NMAE of 0.21% is achieved by the WP using LSTM at a 1-min horizon for non-reschedulable load.

Figure 9:

Figure 10:

Figure 11:

Figure 12:

Table 3 shows that for reschedulable load, the MO achieves the lowest NMAE for all algorithm types and time horizons. At very short-term prediction, deep learning models outperform tree-based models. It suggests that the MO power usage pattern at short-term time scales shows strong temporal dependence, so it is better captured by temporal-based models such as RNNs. However, at a longer horizon (1-day), the temporal dependency decreases, so that non-temporal models, such as RF or LGBM, outperform temporal-based models.

For the non-reschedulable load, no single appliance achieves the lowest NMAE across all scenarios. For a short-term time horizon, the deep-learning models perform better than tree-based. Interestingly, at a 1-day time horizon, the tree-based model (LGBM) shows the lowest NMAE for LG power prediction. For a stable daily power pattern, such as lighting, a temporal-based model might not exploit it.

Overall comparison in Table 3 indicates that no single prediction algorithm shows superiority across the different appliances and time horizons. Deep learning models perform better for short-term forecasting and for appliances with complex temporal dependencies. Tree-based models perform better at longer time horizons or for appliances with stable usage patterns.

Figures 13–16 depict the NRMSE of individual appliance power predictions for 1-min, 15-min, 1-hr, and 1-day time horizons. Table 4 compares the lowest NRMSEs for individual-appliance power predictions across different scenarios, where the lowest NRMSE of 2.45% is achieved by the MO using an RNN at a 1-min horizon for reschedulable load, and the lowest NRMSE of 3.14% is achieved by the RFG using GRU at a 1-min horizon for non-reschedulable load.

Figure 13:

Figure 14:

Figure 15:

Figure 16:

For reschedulable load, the MO achieves the lowest NRMSE for all scenarios. At very short-term time horizons (1-min and 15-min), the tree-based model shows slight superiority over the deep learning model. Oppositely, at 1-hr and 1-day time horizons, the deep learning model shows slight superiority over tree-based models.

For non-reschedulable load, no single appliance achieves the lowest NRMSE across all scenarios. The tree-based models outperform deep learning at longer time horizons.

Overall comparison in Table 4 indicates that tree-based models perform better at predicting non-reschedulable loads, while deep learning models perform better for reschedulable loads.

Figures 17 and 18 show the NMAE and NRMSE of multi-horizon average appliance power predictions, respectively. Figure 17 shows that the RF achieves the lowest NMAE of 0.69% for a 1-min time horizon, the RNN achieves the lowest NMAE of 3.06% for a 15-min time horizon, the GRU achieves the lowest NMAE of 2.88% for a 1-hr time horizon, and the RF achieves the lowest NMAE of 3.84% for a 1-day time horizon.

Figure 17:

Figure 18:

Figure 18 shows that the LGBM achieves the lowest NRMSE of 4.22% for a 1-min time horizon, the LSTM achieves the lowest NRMSE of 9.09% for a 15-min time horizon, the GRU achieves the lowest NRMSE of 9.03% for a 1-hr time horizon, and the RF achieves the lowest NRMSE of 9.31% for a 1-day time horizon. The above results suggest that the tree-based models outperform the deep learning models for multi-horizon average appliance power prediction at the very short (1-min) and long (1-day) time horizons.

Figures 17 and 18 demonstrate that the NMAE and NRMSE increase when the time horizon increases for all prediction techniques. It suggests that prediction accuracy decreases as the forecasting interval increases. Both tree-based and deep-learning models face increasing difficulty in capturing the underlying appliance behavior.

Evaluation of total power prediction

c.

Figures 19–22 depict the NMAE of total power predictions for 1-min, 15-min, 1-hr, and 1-day time horizons. In each figure, the NMAEs for six methods are compared across four different feature groups. The comparison results are summarized in Table 5. Table 5 shows that the lowest NMAE of 2.03% is achieved by the Feature Group-C using an LGBM or XGB at a 1-min horizon for the tree-based algorithm, and the lowest NMAE of 2.53% is achieved by the Feature Group-C using GRU at a 1-min horizon for the deep learning algorithm.

Figure 19:

Figure 20:

Figure 21:

Figure 22:

For the tree-based algorithm, Feature Group-C achieves the lowest NMAE across all time horizons, where LGBM is superior at 1-min and 15-min time horizons, XGB is superior at 1-min and 1-hr, and RF is superior at 1-day. It suggests that the Feature Group-C, which combines all appliance power consumption and total power as the input features of the prediction models, effectively reduces the NMAE.

For the deep learning algorithm, Feature Group-C achieves the lowest NMAE at 1-min and 1-hr time horizons, where GRU is superior, and LSTM is superior at 1-hr and 1-day time horizons. Feature Group-A achieves the lowest NMAE at 1-day time horizon using the LSTM. Again, the Feature Group-C effectively reduces the NMAE, especially for very short-term prediction.

Figures 23–26 depict the NRMSE of total power predictions for 1-min, 15-min, 1-hr, and 1-day time horizons. In each figure, the NRMSEs for six methods are compared across four different feature groups. The comparison results are summarized in Table 6, where the lowest NRMSE of 4.61% is achieved by the Feature Group-C using an LGBM at a 1-min horizon for the tree-based algorithm, and the lowest NRMSE of 4.91% is achieved by the Feature Group-C using GRU at a 1-min horizon for the deep learning algorithm.

Figure 23:

Figure 24:

Figure 25:

Figure 26:

For the tree-based algorithm, Feature Group-C achieves the lowest NRMSE at very short-term time horizons, where LGBM is superior at 1-min and 15-min time horizons, XGB is superior at 1-hr. Feature Group B and D achieve the lowest NRMSE at 1-day using the RF. It suggests that the Feature Group-C effectively reduces the NRMSE, especially for very short-term prediction.

For the deep learning algorithm, Feature Group-C achieves the lowest NRMSE at very short-term time horizons, where GRU is superior at a 1-min time horizon, LSTM at a 15-min time horizon. Feature Group-B achieves the lowest NRMSE at 1-hr time horizon using the RNN. Feature Group-A achieves the lowest NRMSE at 1-day using the RNN. Similarly, before that, the Feature Group-C effectively reduces the NRMSE, especially for very short-term prediction.

Figures 27 and 28 show the NMAE and NRMSE of multi-horizon total power predictions, respectively. Figure 27 shows that the Feature Group-C achieves the lowest NMAE of 2.32% for a 1-min time horizon, the Feature Group-B achieves the lowest NMAE of 5.20% for a 15-min time horizon, the Feature Group-C achieves the lowest NMAE of 6.03% for a 1-hr time horizon, and the Feature Group-A achieves the lowest NMAE of 7.91% for a 1-day time horizon.

Figure 27:

Figure 28:

Figure 28 shows that Feature Group-C achieves the lowest NRMSE of 4.83% for a 1-min time horizon, the lowest NRMSE of 8.31% for a 15-min time horizon, the lowest NRMSE of 9.64% for a 1-hr time horizon, and Feature Group-D achieves the lowest NRMSE of 11.31% for a 1-day time horizon. The results suggest that the Feature Group-C is an effective feature in multi-horizon average total power prediction, especially for the very short-term forecasting.

Figures 27 and 28 demonstrate that the NMAE and NRMSE increase when the time horizon increases, regardless of feature group. A longer horizon makes it harder for the algorithm to capture the model, even with several features adopted.

Evaluation summary

d.

From previous discussions, several findings can be summarized as follows:

1. Individual appliance power prediction:

   1. The NMAE comparison shows that deep learning models are advantageous for short-term forecasting and appliances with complex temporal behavior. In contrast, tree-based models are more effective for longer horizons or appliances with stable usage patterns.

   2. The NRMSE comparisons show that tree-based models are more effective for non-reschedulable loads, while deep-learning models provide advantages for forecasting of reschedulable loads.

   3. The NMAE and NRMSE increase with the time horizon, regardless of the prediction algorithm. It indicates that the prediction algorithm becomes increasingly challenging to capture the underlying appliance behavior over a longer horizon.

2. Total power prediction:

   1. Feature Group-C, which combines all individual appliance power consumption with total power, achieves the lowest NMAE and NRMSE across most horizons for both tree-based and deep-learning models, particularly for very short-term forecasts.

   2. Among tree-based methods, no single model performs across multiple horizons. In deep learning, GRUs and LSTMs dominate short-term horizons.

   3. The NMAE and NRMSE increase when the time horizon increases, regardless of feature group. It indicates that the algorithm is more complicated to capture the model, even with several features adopted.

3. The appliance-specific, horizon-dependent model selection strategy is essential for optimizing prediction accuracy in HEMS.

4. Incorporating all individual appliance power measurements, along with total power, as model inputs enhances total power prediction accuracy.