Accurate multi-horizon photovoltaic (PV) power forecasting is essential for grid stability, energy scheduling, and the efficient operation of PVbatterygrid systems. However, existing models often fail to jointly capture short-term variability and long-term temporal patterns, while lacking unified multi-horizon capability and interpretability. To address these limitations, this study proposes a novel hybrid deep learning framework that integrates a Temporal Fusion Transformer (TFT) and N-BEATS within an attention-based fusion architecture. The proposed model leverages TFT to learn short-term temporal dependencies and dynamic feature importance, while N-BEATS captures long-term trend and seasonal components through interpretable basis decomposition. An adaptive attention mechanism dynamically fuses both branches, enabling balanced multi-scale learning. The framework simultaneously generates forecasts for 30-minute, 2-hour, 6-hour, 12-hour, and 24-hour horizons within a single unified pipeline. The model is evaluated using multi-year real-world datasets from the Desert Knowledge Australia Solar Centre (DKASC). Experimental results demonstrate superior performance over statistical, machine learning, and deep learning baselines. The proposed model achieves RMSE values of 42.35, 46.27, 52.29, and 48.11 across the 20162017, 2023, 2024, and 20252026 datasets, respectively, corresponding to a 3540% reduction compared to the best baseline (TFT). Additionally, MAPE remains below 5% across all datasets, indicating high predictive accuracy and stability. The model also provides interpretability through attention visualization and trendseasonality decomposition, and is deployed on a cloud-based platform for real-time forecasting. These results demonstrate that the proposed framework is robust, interpretable, and suitable for practical renewable energy applications.
Figures & tables
Study
Method
Forecast Horizon
Dataset
Key Limitation
Kim et al. Kim et al. (2026)
MOS + Random Forest
Short-term
Rooftop PV (Korea)
Limited temporal modeling
Salehin et al. Salehin (2026)
KNN–SVM
Short-term
NASA POWER + local data
Requires manual feature engineering
Yousuf et al. Yousuf et al. (2026)
XGBoost + LSTM
Multi-step
UNISOLAR (42 PV sites)
Strong dependence on lag features
Fennane et al. Fennane et al. (2026)
LSTM + TCN
Short-term
PVGIS (Morocco)
Focus on short-term forecasting
Cisse et al. Cisse et al. (2026)
TCN + GRU + RF
Multi-horizon
Solar plant (Australia)
Ensemble complexity
Han et al. Han et al. (2025)
ICEEMDAN + HFNN
Multi-horizon
DKASC, Solar I
Requires signal decomposition
Table 1: Summary of representative studies in photovoltaic power forecasting
Figure 1: Proposed hybrid multi-horizon photovoltaic (PV) forecasting framework. The pipeline includes data preprocessing, dual-branch modeling using TFT and N-BEATS, attention-based fusion, and multi-horizon prediction with evaluation against baseline models.
Dataset
Time Coverage
DKASC-2016-17
2016–2017
DKASC-2023
2023
DKASC-2024
2024
DKASC-2025-26
2025–2026
Table 2: Time coverage of the multi-year photovoltaic datasets used in this study.
Feature
Description
Active Power
Electrical power generated by the PV system (target variable)
Global Horizontal Irradiance
Solar radiation received on a horizontal surface
Diffuse Horizontal Irradiance
Scattered solar radiation component
Ambient Temperature
Air temperature measured at the weather station
Relative Humidity
Percentage of atmospheric moisture
Wind Speed
Wind speed measured at the weather station
Table 3: Key photovoltaic and meteorological variables used in the forecasting model.
Horizon
Forecast Steps
Purpose
30 Minutes
6 steps
Real-time grid stability
2 Hours
24 steps
Short-term dispatching
6 Hours
72 steps
Intraday planning
12 Hours
144 steps
Day-ahead preparation
24 Hours
288 steps
Long-term storage management
Table 4: Forecasting horizons, steps, and purposes in photovoltaic power forecasting
Dataset
Train Samples (70%)
Validation Samples (15%)
Test Samples (15%)
DKASC-2016-17
147743
31660
31659
DKASC-2023
71331
15286
15285
DKASC-2024
71331
15286
15285
DKASC-2025-26
73770
15808
15808
Table 5: Chronological data splitting and sample sizes for each photovoltaic dataset.
Figure 2: Data preprocessing pipeline applied to the photovoltaic dataset for PV power forecasting.
Figure 3: Proposed Multi horizon Model
Figure 4: Attention-Based Fusion Mechanism
Horizon
Forecast Steps
Purpose
30 Minutes
6 steps
Real-time grid stability
2 Hours
24 steps
Short-term dispatching
6 Hours
72 steps
Intraday planning
12 Hours
144 steps
Day-ahead preparation
24 Hours
288 steps
Long-term storage management
Table 6: Forecast horizons and their purposes
Parameter
Value
Optimizer
Adam
Learning Rate (Phase 1)
1×10−3
Learning Rate (Phase 2)
1×10−4
Weight Decay
1×10−5
Scheduler
ReduceLROnPlateau
Table 7: Optimization Parameters
Hyperparameter
Value
Description
Sequence Length
144
Historical time steps used as input
Forecast Horizons
6, 24, 72, 144, 288
Prediction steps for multiple time scales
Batch Size
64
Samples per training batch
Hidden Dimension
64
Size of neural network representations
LSTM Layers
1
Number of LSTM layers in TFT branch
Attention Heads
4
Multi-head attention components
Table 8: Hyperparameter Settings for the Proposed Hybrid Forecasting Model
Dataset
RMSE
MAE
MAPE
DKASC-2016-2017
42.35
35.41
2.88
DKASC-2023
46.27
38.00
3.29
DKASC-2024
52.29
40.08
4.02
DKASC-2025-2026
48.11
39.61
3.91
Table 9: Cross-dataset forecasting performance of the proposed model
Model
2016–2017
2023
2024
2025–2026
RMSE
MAE
MAPE
RMSE
MAE
MAPE
RMSE
MAE
MAPE
RMSE
MAE
MAPE
ARIMA
87.12
73.84
12.31
87.12
73.84
12.31
89.12
71.84
14.31
84.12
77.84
14.31
SARIMA
94.67
78.43
14.56
94.67
78.43
14.56
91.67
82.43
12.56
91.67
82.43
11.56
Random Forest
91.35
84.90
17.02
91.35
84.90
17.02
94.35
80.90
13.02
94.35
80.90
13.02
XGBoost
87.91
92.12
18.95
87.91
92.12
18.95
84.91
88.12
21.95
83.91
89.12
21.95
LSTM
82.45
68.32
16.87
82.45
68.32
16.87
78.45
71.32
13.87
79.45
71.32
13.87
Table 10: Baseline and Proposed Model Performance Across Four PV Datasets
Model Variant
2016–2017
2023
2024
2025–2026
RMSE
MAE
MAPE
RMSE
MAE
MAPE
RMSE
MAE
MAPE
RMSE
MAE
MAPE
TFT only
65.41
58.91
11.41
65.41
58.91
11.41
68.41
55.91
13.41
68.41
54.91
13.41
N-BEATS only
73.67
65.81
19.45
73.67
65.81
19.45
70.67
68.81
16.45
70.67
68.81
16.45
Hybrid (without attention)
68.91
71.25
13.84
68.91
71.25
13.84
65.91
74.25
11.84
71.91
68.25
16.84
Hybrid (full model)
42.35
35.41
2.88
46.27
38.00
3.29
52.29
40.08
4.02
48.11
39.61
3.91
Table 11: Ablation Study of the Proposed Hybrid Model Across Four PV Datasets
Figure 5: Distribution comparison of actual and predicted values across different datasets
Figure 6: SHAP-based global feature importance analysis across four photovoltaic datasets. The figures illustrate the most influential features contributing to model predictions under varying temporal conditions.
Figure 7: The real world application of proposed model, showing the interface.
Solar power forecasting predicts photovoltaic generation and irradiance from weather, imagery, plant data, and recent output to support reliable grid and energy decisions.
School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China · University of Southern Queensland, Toowoomba 487-535, Australia · School of Information Engineering, Tianjin University of Commerce, Tianjin 300134, China +3