Photovoltaic Decomposition Method Based on Multi-Scale Modeling and Multi-Feature Fusion

Abstract

Deep learning-based Non-Intrusive Load Monitoring (NILM) methods have been widely applied to residential load identification. However, photovoltaic (PV) loads exhibit strong non-stationarity, high dependence on weather conditions, and strong coupling with multi-source data, which limit the accuracy and generalization of existing models. To address these challenges, this paper proposes a multi-scale and multi-feature fusion framework for PV disaggregation, consisting of three modules: Multi-Scale Time Series Decomposition (MTD), Multi-Feature Fusion (MFF), and Temporal Attention Decomposition (TAD). These modules jointly capture short-term fluctuations, long-term trends, and deep dependencies across multi-source features. Experiments were conducted on real residential datasets from southern China. Results show that, compared with representative baselines such as SGN-Conv and MAT-Conv, the proposed method reduces MAE by over 60% and SAE by nearly 70% for some users, and it achieves more than 45% error reduction in cross-user tests. These findings demonstrate that the proposed approach significantly enhances both accuracy and generalization in PV load disaggregation.

1. Introduction

With the continuous rise in demand for energy efficiency [1] and intelligent management [2,3], distributed photovoltaic (PV) systems have gradually become an integral part of residential energy systems. This trend imposes higher demands on distribution-side energy management, particularly requiring high-precision monitoring and optimization of distributed energy resources without additional hardware investments. Non-Intrusive Load Monitoring (NILM), a technology that relies solely on smart meter data, offers a viable solution to this challenge [4,5]. This study hypothesizes that multi-scale modeling and multi-feature fusion can effectively capture the non-stationary characteristics and multi-source feature dependencies of PV loads, thereby enhancing decomposition accuracy and generalization. Based on this, the research objective is to design and validate a PV decomposition framework suitable for residential scenarios, while also exploring its scalability across different users and contexts. As shown in

Table 1. Symbol Name and Meaning.

Symbol	Meaning	Symbol	Meaning
NILM	Non-Intrusive Load Monitoring	$y_i$	The compressed representation of the $x_i$
PV	Photovoltaic	${\alpha }_i$	weights
MTD	Multi-scale Time-series Decomposition	$\odot$	elemental multiplication
MFF	Multi-feature Fusion	$U$	Multi-scale fusion feature representation
TAD	Temporal Attention Decomposition	$U_{time}$	Convergence characteristics of the time dimension
MAE	Mean Absolute Error	$U_{channel}$	Convergence characteristics of channel dimensions
MSE	Mean Squared Error	Z	Dual-channel fusion features
SAE	Symmetric Absolute Error	${\mathrm{h}}_{\mathrm{c}}$	Time representation for each channel
CNN	Convolutional Neural Network	W	mapping matrix
LSTM	Long Short-Term Memory	D	Feature Dimension
GRU	Gated Recurrent Unit	d	Model Embedding Dimension
UAS	Universal Adaptive Stabilization	$W_Q,W_K,W_V$	Learning parameters for Q, K, V
CMM	Consumer Mixture Model	$C_t$	context-enhanced representation
FNN	Feedforward Neural Network	$E_t$	position code
ReLu	Rectified Linear Unit	$b_0$	bias value
X	Inputs to the model	${\widehat{y}}^{PV}$	Projected value of PV
T	Length of time series	$w_0$	projection matrix
C	Number of channels	${\mathcal{L}}_{PV}$	loss function
k	Number of down-sample	Softmax (·)	activation function
$x_i$	Subsequent sequences of multiscale decompositions	MeanPool (·)	Average pooling

, for clarity and consistency, a comprehensive table of symbols is provided to assist the reader in following the technical presentation.

Traditional load decomposition methods have made progress in identifying household appliances (e.g., refrigerators, washing machines) [6,7], but decomposing PV output remains more challenging. Driven by solar irradiation and diurnal variations, PV signals exhibit high volatility and nonlinear characteristics while often overlapping with other loads, significantly increasing task complexity [8,9]. Existing research has attempted improvements through deep Auto-Transformer models [10], enhanced clustering methods [11], or ensemble learning approaches [12]. Nevertheless, three major limitations remain: First, single-timescale approaches struggle to balance short-term fluctuations with long-term trends; Second, insufficient utilization of meteorological data prevents full exploitation of cross-dimensional feature interactions; Third, reliance on complex models or prior knowledge restricts scalability across diverse scenarios. Therefore, there is an urgent need for an innovative framework that simultaneously incorporates multi-scale modeling and multi-source feature fusion to address these limitations.

To tackle these challenges, this paper proposes a multi-scale and multi-feature fusion photovoltaic decomposition framework specifically tailored for residential scenarios. The framework consists of three modules: First, Multi-Scale Time Series Decomposition (MTD), which captures both local fluctuations and long-term trends through hierarchical down-sampling and fusion; Second, Multi-Feature Fusion (MFF), which models semantic interactions between load and meteorological variables across temporal and channel dimensions; Third, Temporal Attention Decomposition (TAD), which employs attention mechanisms to identify key time segments strongly correlated with PV output, thereby enhancing robustness under low-sampling conditions. Compared with existing approaches, the proposed framework achieves joint modeling across temporal scales and feature dimensions, while avoiding reliance on high-frequency sampling and complex prior knowledge, thus substantially improving accuracy and generalization. It should be noted that the dataset used in this study originates from residential load data, and further validation is required for commercial or industrial applications.

The contributions of this paper are threefold:

• Proposing and validating a multi-scale and multi-feature fusion PV decomposition framework based on the hypothesis that joint modeling across temporal scales and feature dimensions enhances accuracy and generalization;
• Organically integrating multi-scale decomposition, feature interaction modeling, and attention mechanisms in the framework design, which differs from traditional NILM methods that rely on single-scale or shallow fusion approaches;
• Demonstrating significant improvements over SGN-Conv and MAT-Conv on real residential user data from southern China, while maintaining robust performance across different user experiments. It is important to note that the design and validation of this method are based on residential scenarios, and additional testing is necessary for commercial or industrial PV systems.

2. Related Work

In recent years, deep learning has demonstrated remarkable capabilities in feature extraction and sequence modeling, significantly advancing the performance of NILM. Particularly under low-frequency sampling conditions, traditional statistical methods have been progressively replaced by end-to-end neural network architectures such as Long Short-Term Memory Networks (LSTM) [13], Convolutional Neural Networks (CNN) [14], and Transformer-based models [15,16]. These approaches automatically capture the nonlinear dynamics and long-range dependencies inherent in aggregate load data, achieving relatively good results on various public NILM datasets. Figure 1 shows that the number of searches for the keyword “NILM” has continued to grow over the past six years, reflecting the research enthusiasm and attention in this field. With the increasing penetration of distributed PV systems in residential settings, researchers have begun integrating PV generation into the NILM framework. This integration enables the estimation of PV output directly from the total load curve recorded by smart meters, eliminating the need for additional sensors. To provide a clearer overview, recent studies are systematically compared in terms of their methodological characteristics and input features, with the results summarized in

Table 2. Comparative Analysis of Existing Methods, “✕” indicates that the method is not used, and “✓” indicates that the method is used.

Methods	Weather Data Fusion	Multi-Scale Modeling	Dual-Channel Fusion
[13,15]	✕	✕	✕
[6]	✓	✕	✕
[14]	✓	✓	✕
Ours	✓	✓	✓

.

Beyond directly decomposing PV data from the total load curve, some studies have attempted modeling and control parameter estimation at the inverter level. For instance, ref. [17] employed Gaussian mixture models and curve fitting to reconstruct reactive power parameters, thereby avoiding reliance on distribution network topology. However, this method is highly sensitive to sunrise and sunset times, resulting in significantly diminished performance in distributed scenarios with multiple parallel inverters. In contrast, the proposed method requires no additional inverter-level modeling; it achieves stable decomposition solely based on aggregate load and meteorological data, thereby overcoming the applicability limitations of such approaches in distributed scenarios.

Similarly, ref. [18] combined multiscale modeling with a meteorological attention mechanism to enhance dynamic PV modeling capabilities. However, this approach heavily relies on high-quality meteorological inputs and experiences significant performance degradation under data-deficient or noisy conditions. It also lacks robustness in non-stationary load scenarios involving energy storage and electric vehicles. In contrast, our method employs a Multi-Feature Fusion (MFF) module for deep interactive modeling of electrical and meteorological features, maintaining robust performance even with low-quality meteorological data and thereby offering greater practicality. Ref. [19] proposed a Transformer–CNN hybrid architecture incorporating a “time–application attention” mechanism to capture complex device dependencies, primarily targeting industrial scenarios. Its multi-scale attention offers some reference value for non-stationary signals, but its transferability is limited in residential PV decomposition due to diverse user loads and energy storage interventions. In contrast, the proposed method’s TAD module is specifically designed for residential PV scenarios. It focuses on PV-related time segments under low-sampling conditions, better aligning with residential user needs.

With advances in deep learning, some studies have attempted to reduce dependence on meteorological data. For instance, ref. [20] proposed a real-time framework based on Universal Adaptive Stabilization (UAS), estimating PV and residential loads solely from net load data with rapid convergence. However, this method requires manual parameter tuning and performs poorly under extreme weather conditions. In contrast, the proposed method automatically models the relationship between meteorological data and load through Multi-Scale Time Series Decomposition (MTD) and the MFF module, eliminating the need for complex parameter tuning and demonstrating stable performance across diverse weather conditions.

Ref. [21] introduced a “month-to-hour mapping” strategy, employing Gaussian mixture models and reference PV curves, which is suitable for low-resolution data. However, it relies heavily on reference curves and is prone to errors in cross-climate or energy-storage scenarios. Ref. [22] proposed the Consumer Mixture Model (CMM), which estimates PV output using neighboring users’ nighttime curves and supports multi-user aggregation. However, it relies on region-specific characteristics, leading to significant errors in cases of energy storage discharge or cross-regional deployment. In contrast, the proposed method does not depend on specific reference curves or regional features. Instead, it automatically adapts to different scenarios through multiscale decomposition and cross-feature interaction modeling, thereby enhancing its generalization capability.

In summary, existing NILM approaches for PV disaggregation face several critical limitations. First, most models are confined to a single temporal resolution, making it difficult to simultaneously capture sudden variations and long-term trends in PV output. Second, the integration of meteorological and electrical features is often superficial, lacking structured mechanisms for multi-dimensional interaction. Third, many methods involve complex architectures or require prior domain knowledge, which compromises generalization and limits real-world applicability.

To address these challenges, this paper proposes a deep learning-based PV disaggregation framework that integrates multi-scale temporal modeling with semantic feature fusion. Specifically, the method employs a down-sampling-based strategy to construct multi-resolution temporal sequences and introduces a dual-channel architecture that models interactions between electrical loads and meteorological variables across temporal and channel dimensions. This design enhances the model’s capability to represent complex PV dynamics while improving its adaptability to heterogeneous environmental conditions.

3. Method

As illustrated in Figure 2, this paper proposes a non-intrusive modeling framework for PV load disaggregation, consisting of three core modules: Multi-scale Time-series Decomposition (MTD), Multi-feature Fusion (MFF), and Temporal Attention-based Decomposition (TAD). The MTD module takes the aggregated load and associated environmental variables $X\in R^{T\times C}$ as input, where T denotes the sequence length and C denotes the number of features, extracts multi-scale features via down-sampling, and integrates them through an attention mechanism to produce a unified representation $U_{fused}$. The MFF module then models temporal and channel-level dependencies to enhance the representation, yielding feature sequence Z. Finally, the TAD module applies a Transformer-based attention mechanism to capture contextual information and estimate the PV load sequence ${\widehat{y}}_t^{PV}$. These three modules work in a coordinated manner, enabling accurate modeling and separation of PV components from the total load. Implementation details of each module are elaborated in the following sections.

3.1. Multi-Scale Time-Series Decomposition Module

To capture dynamic patterns across different time scales in photovoltaic load curves, this study proposes an MTD module. Through down-sampling operations, X is divided into multiple sub-sequences $\{x_1,x_2,\cdots ,x_k\}$ with varying temporal resolutions, where k denotes the number of down-sampling layers. This approach simultaneously preserves rapid fluctuations and long-term trends, avoiding the omission of critical patterns common in traditional single-scale methods.

To effectively fuse features across scales, MTD employs an attention-guided feature fusion mechanism. First, each sub-sequence $x_i$ undergoes average pooling to yield a compressed context vector $y_i$ representing the core information at that scale. Next, all $\{y_1,y_2,\cdots ,y_k\}$ are concatenated and fed into a fully connected layer. A Softmax function generates attention weights ${\alpha }_i$ to measure the importance of each scale. Finally, MTD obtains the fused temporal feature representation through weighted summation:

$${\mathrm{U}}_{\mathrm{fused}}={\displaystyle \sum _{i=1}^k{\alpha }_i{x}_i}$$

(1)

Here, $a_i$ represents the weight learned by the attention mechanism, determining each scale’s contribution to the final representation. The resulting $U_{fused}$ simultaneously captures local fluctuations and global trends, providing more robust inputs for subsequent modeling.

Compared to existing NILM methods, this study innovates by “first performing explicit multi-scale decomposition, then applying attention fusion.” Traditional multi-scale approaches typically rely on fixed rules, lacking adaptability; attention mechanisms often operate directly on raw sequences, struggling to fully leverage multi-scale features. MTD combines both, enabling the model to dynamically select scales based on varying household and meteorological conditions, thereby achieving greater robustness in photovoltaic decomposition tasks.

3.2. Multi-Feature Fusion Module

To model the dynamic dependencies between load and meteorological variables, this study proposes a MFF module. Taking $U_{fused}$ as input, this module employs a Gated Recurrent Unit (GRU) network to capture the temporal evolution of each variable while utilizing a channel attention mechanism to characterize interactions between variables. This achieves deep fusion of multi-source features, enhancing feature representation capabilities.

In the temporal modeling component, each input variable sequence $U\left[:c\right]$ is fed into a GRU with shared parameters to extract its temporal dependency feature $h_c$. Representations of all variables are then stacked into a unified tensor $U_{time}$, preserving the dynamic patterns of each variable over time. This enables the model to recognize synergistic effects on PV output, such as high temperature and strong irradiation.

To further characterize variable interactions, MFF introduces a channel attention mechanism at each time step. This mechanism assigns weights $a_t$ to each variable, quantifying its importance in the global representation. Based on the attention distribution, a channel-weighted feature representation $U_{channel}$ is obtained. Additionally, a learnable gating coefficient β ∈ [0,1] is introduced to regulate the contribution of interaction features, preventing certain redundant variables from amplifying noise.

During the fusion stage, temporal features $U_{time}$ and channel features $U_{channel}$ are integrated according to their weights, while residual connections preserve low-level features from the original input U. The resulting fused representation Z incorporates both single-variable dynamic information and explicitly models contextual dependencies among multiple variables.

$$Z=U+U_{time}+\beta \cdot U_{channel}$$

(2)

Compared to existing NILM methods, MFF’s novelty lies in its “dual-layer modeling of time dependency and channel interaction.” Traditional approaches often model only the load curve itself, neglecting the influence of meteorological factors, or treat meteorological variables merely as external inputs without deeply modeling their semantic relationship with load. MFF, however, combines temporal modeling with channel modeling, enabling load and meteorological features to interact at a deep semantic level. This significantly enhances the model’s interpretability and robustness in photovoltaic decomposition tasks.

3.3. Temporal Attention-Based Decomposition Module

After obtaining the fused feature sequence $Z\in R^{T\times C}$, the TAD module further leverages a Transformer-based structure to model temporal dependencies. By incorporating learnable positional encodings and a multi-head self-attention mechanism, this module extracts contextual features related to PV generation, enabling accurate PV load disaggregation.

The module first applies a linear projection to the input sequence Z, followed by the addition of learnable positional embeddings to provide temporal order awareness. This process is formulated as:

$$X_t=ZW+E_t$$

(3)

where $W\in R^{D\times d}$ is the embedding matrix that maps the fused features into the Transformer modeling space, and $E_t$ is the positional encoding at time step t. Here, D denotes the feature dimension, while d represents the model embedding dimension within the transformer. The resulting embedded sequence $X_t\in R^{T\times d}$ is then fed into the temporal modeling module.

To capture dependencies among time steps, a multi-head self-attention mechanism is employed. This mechanism enables each time step to interact with all others through global attention. The attention computation is given by:

$${\alpha }_{\mathrm{t}}=\mathrm{softmax}({\displaystyle \frac{Q{K}^{{\rm T}}}{\sqrt{d}}})$$

(4)

$${\mathrm{C}}_t={\alpha }_t{V}_t$$

(5)

where $Q=X_t{W}_Q, K=X_t{W}_K, V=X_t{W}_V$ represent the query, key, and value matrices, respectively, with learnable parameters $W_Q, W_K, W_V\in R^{d\times d}$. ${\alpha }_t$ denotes the attention weight, and $C_t$ is its context-enhanced representation.

Subsequently, a time-distributed feedforward neural network maps each context vector $C_t$ to its corresponding PV load estimate:

$${\widehat{y}}_t^{PV}=Relu(w_0{C}_t+b_0)$$

(6)

where $W_0\in R^{d\times 1}$, $b_0\in R$, and ReLu(·) is the activation function. The final output sequence ${\widehat{y}}^{PV}\in R^T$ represents the disaggregated PV load component.

To guide the model to focus more on the structural changes of the PV load, especially during periods of significant magnitude fluctuation, a gradient-sensitive time-weighted loss function is introduced:

$${\mathcal{L}}_{PV}={\displaystyle \frac{1}{T}}{\displaystyle \sum _t^T{w}_t{({\widehat{y}}_t^{PV}-y_t^{PV})}^2}$$

(7)

where the weight $w_t$ is determined based on the local gradient of the ground-truth PV load. When $\left|y_t-y_{t-1}\right|>\theta$, a larger weight is assigned to enhance the model’s focus on abrupt changes.

By introducing attention-based temporal modeling, this module enables the model to utilize global temporal context in the disaggregation process. It effectively overcomes the equal-weight limitations of traditional methods and enhances the model’s ability to capture time-sensitive structures, making it a core component of the NILM disaggregation pipeline.

4. Experiment

4.1. Dataset Description

The experiments in this paper are conducted using a real-world residential energy consumption dataset collected from a city in southern China. This dataset includes net load and distributed PV generation data from four residential users, along with corresponding meteorological information such as temperature and solar irradiance, all aligned with timestamp records. To further evaluate cross-user generalization, we included two additional households, not used in training, as independent test subjects in our experiments. The data is sampled at an hourly resolution, spanning nearly two years, resulting in approximately 17,000 samples per user. These sequences are then split into training, validation, and test sets using a 7:2:1 ratio. It should be explicitly noted that the dataset contains a limited number of users and is sourced exclusively from residential scenarios within a single city. Consequently, the conclusions drawn in this paper are subject to limitations in terms of geographical scope and scale.

4.2. Parameter and Hardware Configuration

For model configuration, the input sequence length is set to 720, with predictions made at the same temporal resolution. The overall architecture comprises three layers of multi-scale fusion, each with a scale width of 2; the patch size is 16, and the dropout rate is 0.1. The Adam optimizer is employed with an initial learning rate of 0.001. Training is conducted for up to 100 epochs with early stopping enabled, using a batch size of 128. To enhance robustness under low-output conditions, the loss function incorporates a zero-output penalty term in addition to the standard MSE, penalizing non-zero predictions when the actual PV output approaches zero. All experiments are conducted on a CUDA-accelerated NVIDIA RTX 3090 GPU. It should be noted that the RTX 3090 is used solely to accelerate training and is not a computational requirement for the model, which can also be deployed on an RTX 3060 GPU or a GTX 1080 GPU.

For comparative methods, SGN-Conv [23] and MAT-Conv [24] are implemented according to their original papers and publicly available code. SGN-Conv maintains the same number of convolutional layers and channels as the original paper, using the Adam optimizer with an initial learning rate of 0.001. MAT-Conv’s multi-head attention layers and embedding dimensions also follow the original settings. Across all experiments, the learning rate is uniformly set to 0.001 and the batch size to 128. Both methods employ early stopping to prevent overfitting. Under these unified hardware and training conditions, the configurations ensure a fair comparison between different approaches.

4.3. Evaluation Metrics

To comprehensively evaluate model performance in PV disaggregation, three commonly used error metrics are employed: mean absolute error (MAE), mean squared error (MSE), and symmetric absolute error (SAE).

$$MAE={\displaystyle \frac{1}{T}}{\displaystyle \sum _{t=1}^T|{\widehat{y}}_t-y_t|}$$

(8)

$$MSE={\displaystyle \frac{1}{T}}{\displaystyle \sum _{t=1}^T{({\widehat{y}}_t-y_t)}^2}$$

(9)

$$\mathrm{SAE}={\displaystyle \frac{1}{\mathrm{M}}}{\displaystyle \sum _{\mathrm{m}=1}^{\mathrm{M}}\frac{1}{\mathrm{T}}}|{\displaystyle \sum _{\mathrm{t}=1}^T{y}_t}-{\displaystyle \sum _{t=1}^T{y}_t}|$$

(10)

where ${\widehat{y}}_t$ denotes the disaggregated PV output at time step t, $y_t$ denotes the ground-truth PV generation, M represents the batch size, and T represents the number of samples contained in each batch.

4.4. Main Experimental Results

Table 3. The experimental results are shown below, bold text indicates the best results.

User	Model	MAE	MSE	SAE
User 1	SGN-Conv	0.4754	0.4646	0.2350
	MAT-Conv	0.4420	0.4711	0.1968
	Ours	0.1819	0.1007	0.0781
User 2	SGN-Conv	0.2112	0.0913	0.1078
	MAT-Conv	0.2087	0.0981	0.0622
	Ours	0.0883	0.0239	0.0191
User 3	SGN-Conv	0.3644	0.2643	0.1888
	MAT-Conv	0.2090	0.1233	0.0879
	Ours	0.2318	0.2380	0.0436
User 4	SGN-Conv	0.2332	0.0744	0.0257
	MAT-Conv	0.2719	0.2220	0.2111
	Ours	0.1467	0.0695	0.0664

presents quantitative performance comparisons between the proposed method and two representative existing models SGN-Conv [23] and MAT-Conv [24] for four residential user scenarios. Overall, the proposed method demonstrates significant advantages across most metrics, effectively validating its accuracy and robustness in photovoltaic load decomposition tasks. On User 1 and User 2 datasets, our model achieves MAE values of 0.1819 and 0.0883, respectively—approximately 60% lower than SGN-Conv and 50% lower than MAT-Conv. Visualization results in Figure 3 further demonstrate high consistency between predicted and actual curves, particularly in peak-valley intervals.

In User 3’s experiment, the MAE and MSE of our method were slightly higher than those of MAT-Conv. This discrepancy did not stem from an overall performance degradation but rather from the alignment between signal characteristics and model mechanisms. Specifically, User 3’s photovoltaic output curve exhibited pronounced periodicity and highly stable diurnal variation patterns, with minimal short-term fluctuations (manifested as a low signal-to-noise ratio and small rolling variance). Under such conditions, the MTD module tends to generate redundant components unrelated to the dominant daily cycle during multi-scale decomposition. These components are subsequently amplified in feature fusion and attention mechanisms, leading to overfitting to local noise or minor disturbances and consequently increasing MAE and MSE. Figure 4 illustrates this phenomenon: The upper panel shows that while our model’s predictions (red dashed line) generally align with actual PV output (blue solid line), a sharp overprediction peak emerged around March 19. The residual plot below corroborates this, revealing isolated large errors that primarily drove the increased MAE and MSE values, supporting our hypothesis of localized overfitting anomalies. In contrast, MAT-Conv directly captures subtle differences in temporal neighborhoods through its local attention mechanism, giving it superior performance in “low-fluctuation, strong-periodicity” scenarios. Nevertheless, our method demonstrates stronger preservation of the primary cycle in spectral energy concentration analysis and maintains superiority in SAE metrics, indicating greater robustness in capturing long-term energy trends and suppressing cumulative error propagation. This suggests that even in stable, highly periodic scenarios, our approach retains clear advantages in modeling long-term structures and preserving overall energy patterns.

In user 4’s experiments, our method achieved lower MAE and MSE but slightly underperformed SGN-Conv in SAE. This is mainly because User 4’s PV curves contain a few sharp anomalous peaks. Our framework, emphasizing global trends through multi-scale decomposition, smooths such local spikes, leading to weaker recovery of extreme values and thus higher SAE. By contrast, SGN-Conv’s convolutional structure is more sensitive to short-term fluctuations, enabling better peak capture but at the expense of stability in overall trend modeling. In summary, our method is stronger in balancing global accuracy and long-term error control, while SGN-Conv excels at capturing isolated extreme peaks.

4.5. Ablation Experiment

As shown in

Table 4. Four users’ ablation test results, bold text indicates the best results.

User	Model	MAE	MSE	SAE
User 1	Case 1	0.4736 [0.3980, 0.5358]	0.4669 [0.3679, 0.5758]	0.2257 [0.1806, 0.3376]
	Case 2	0.2589 [0.2238, 0.3056]	0.1775 [0.1358, 0.2091]	0.1037 [0.0883, 0.1311]
	Ours	0.1819 [0.1709, 0.1976]	0.1007 [0.0845, 0.1179]	0.0781 [0.0635, 0.0895]
User 2	Case 1	0.2108	0.0918	0.1074
	Case 2	0.1258	0.0386	0.0468
	Ours	0.0883	0.0239	0.0191
User 3	Case 1	0.2584	0.1534	0.1153
	Case 2	0.1912	0.1083	0.0875
	Ours	0.2318	0.2380	0.0436
User 4	Case 1	0.2047	0.0799	0.0776
	Case 2	0.1693	0.0883	0.1610
	Ours	0.1467	0.0754	0.0664

, we further validated the contribution of each module to overall performance through ablation experiments. Case 1 represents removing meteorological features, while Case 2 denotes removing the MTD module. Overall, both ablation approaches resulted in a significant increase in error, indicating that the proposed modules play a crucial role in enhancing both model accuracy and robustness. Particularly in the results for User 1 and User 2, the complete model significantly outperformed the control group across all three metrics: MAE, MSE, and SAE. Moreover, Case 1 generally exhibited higher errors than Case 2, indicating that meteorological information holds significant value in capturing the non-stationary characteristics of PV output, while the MTD module is indispensable for modeling multi-timescale dynamic variations.

It is worth noting that in User 3’s results, Case 2 outperformed the full model in both MAE and MSE, consistent with the anomalous findings in previous experiments. Further analysis revealed that this user’s PV curve exhibited high stability and low volatility. In such scenarios, multi-scale decomposition may introduce redundant features or noise interference, thereby weakening the model’s ability to capture fine-grained variations. Nevertheless, the full model still delivers the best performance on the SAE metric, indicating its continued advantage in modeling long-term energy trends and controlling cumulative error. Thus, the anomalous results do not negate the MTD module’s utility but rather reveal its potential limitations in specific stable scenarios.

To verify the robustness of the ablation experiment results, we further calculated the 95% confidence intervals of three metrics (MAE, MSE, SAE) for User 1. As shown in the table, both Case 1 and Case 2 exhibited significantly higher errors compared to the complete model, with their confidence interval bounds showing little overlap with those of the full model. For example, the complete model achieved an MAE of 0.1819 [0.1709, 0.1976], which is substantially lower than 0.4736 [0.3980, 0.5358] in Case 1 and 0.2589 [0.2238, 0.3056] in Case 2. Similar trends are observed for the MSE and SAE metrics. These findings demonstrate that the complete model not only outperforms the ablation configurations in terms of average performance but also maintains a stable advantage within the uncertainty range, further confirming the critical role of the MTD module and meteorological features in enhancing both the accuracy and robustness of the model.

4.6. Generalization Experiment

To further validate the model’s generalization capability on unseen users, we evaluated it using newly added User 5 and User 6 as independent test subjects, while retaining the original four users as the training set.

Table 5. Generalization experiment result, bold text indicates the best results.

User	Model	MAE	MSE	SAE
User 5	SGN-Conv	0.3238	0.2522	0.2584
	MAT-Conv	0.3696	0.2147	0.1636
	Ours	0.1778	0.1007	0.1369
User 6	SGN-Conv	0.3817	0.3098	0.2534
	MAT-Conv	0.2316	0.1526	0.1977
	Ours	0.1530	0.0834	0.0724

presents the comparison results against two baseline models. The results demonstrate that the proposed method exhibits significant advantages for both new users. For User 5, our approach achieves MAE and MSE values of 0.1778 and 0.1007, respectively, representing reductions of approximately 45% and 60% compared to SGN-Conv, and 52% and 53% compared to MAT-Conv, significantly improving point prediction accuracy. Simultaneously, it maintains a leading position in the SAE metric, indicating greater robustness in cumulative energy prediction. In the User 6 test, our method achieves MAE and MSE values of 0.1530 and 0.0834, respectively, representing substantial reductions compared to both SGN-Conv and MAT-Conv. The advantage is particularly pronounced in the SAE metric (0.0724, approximately 71% lower than SGN-Conv and 63% lower than MAT-Conv), further validating the model’s reliability in modeling long-term energy trends.

Overall, the proposed method demonstrates consistent performance with previous experimental results on new users, not only improving single-point prediction accuracy but also exhibiting enhanced robustness in long-term energy estimation. This indicates that the multi-scale modeling and feature fusion strategy effectively captures key dynamic characteristics of PV output and possesses a degree of cross-user transferability. Although validation remains confined to the same region, the results provide positive support for the model’s generalizability.

4.7. Sensitivity Analysis and Time Cost Comparison

In this study, we primarily examined two hyperparameters—learning rate and dropout—that most significantly impact model stability and generalization capability. Sensitivity analysis results, as shown in

Table 6. Sensitivity Analysis.

Lr	Dropout	MAE	MSE	SAE
0.001	0	0.1896	0.1048	0.0990
0.001	0.1	0.1907	0.1013	0.0828
0.001	0.3	0.1819	0.1007	0.0781
0.0001	0.1	0.1876	0.1018	0.1659
0.01	0.1	0.5805	0.4347	0.4347

, indicate that learning rate exerts the greatest influence on model performance: when set excessively high (0.01), model performance declines sharply, while remaining stable within the range of 0.001 to 0.0001. In contrast, dropout enhances robustness; moderate settings effectively suppress overfitting while further optimizing metrics. Overall, learning rate is the key factor determining model convergence stability, while appropriately applied dropout plays a supplementary role in boosting generalization.

Regarding computational efficiency, as shown in

Table 7. Time Cost Comparison, bold text indicates the best results.

Model	Time/s
SGN-Conv	8.19
MAT-Conv	15.64
Ours	7.75

, experiments compared SGN-Conv, MAT-Conv, and the proposed method under identical hardware (NVIDIA RTX 3090 GPU), dataset (User 1), and hyperparameter conditions. Using the average training time over the first 100 epochs as the metric, results show that our method achieves significantly higher accuracy while averaging only 7.75 s per epoch. This outperforms SGN-Conv and is substantially lower than MAT-Conv’s 15.64 s, demonstrating a strong efficiency advantage.

5. Conclusions

This paper proposes a specialized NILM method for distributed PV systems. To address the non-stationary characteristics, weather dependency, and strong coupling of PV loads with total residential demand, we design a framework that integrates multiscale decomposition, multi-feature fusion, and temporal attention mechanisms. The proposed framework effectively decomposes PV loads by balancing short-term fluctuations with long-term trends, without requiring additional hardware or domain-specific prior knowledge.

Experimental validation on real residential datasets confirms the effectiveness of the proposed method. The results demonstrate superior performance over representative baseline models across multiple user scenarios, with strong generalization potential observed in cross-user tests. These findings highlight the framework’s significant practical value for residential PV decomposition tasks.

Despite these positive outcomes, several limitations remain: First, the limited number of users in the generalization validation is insufficient to demonstrate robustness at larger scales or across diverse regions; second, the MTD module may introduce redundant representations under stable load patterns, indicating the need for adaptive structural optimization; and third, robustness under extreme weather conditions has not yet been systematically evaluated, while PV volatility during abnormal meteorological events continues to pose a major challenge.

Future research will advance along three directions: (1) expanding to larger-scale and cross-regional datasets to improve generalization across diverse user groups and climatic conditions; (2) integrating with forecasting models and adaptive deployment strategies within smart grid management systems to support more complex energy optimization scenarios; and (3) conducting systematic studies on resilience under extreme weather and heterogeneous loads (e.g., energy storage, electric vehicles), while evaluating scalability for commercial and industrial PV systems. These efforts aim to further enhance both the scalability and the practical value of the proposed NILM framework in real-world distributed energy environments.