Predicting Photovoltaic Power Output Using LSTM: A Comparative

Study Using both Historical and Climate Data

Fereshteh Jafari

1,2 a

, Joseph Moerschell

1 b

and Kaspar Riesen

2 c

Institute of System Engineering, University of Applied Sciences and Arts Western Switzerland, Sion, Switzerland

Institute of Computer Science, University of Bern, CH-3012 Bern, Switzerland

Keywords:

LSTM, Meteorological Data Integration, Photovoltaic Power Prediction, Prediction Horizon Assessment,

Temporal Sliding Window Analysis.

Abstract:

Accurate photovoltaic (PV) power output prediction is important for efﬁcient energy management in solar

power systems. This study explores the beneﬁts and limitations of Long Short-Term Memory (LSTM) net-

works in predicting PV power using three distinct approaches, namely using historical PV power data, climate

data, and a combination of both, all with timestamps. The performance of these methods is evaluated across

different prediction horizons of 10, 30, and 50 minutes ahead. Additionally, the impact of the sliding window

size, representing the amount of past data used for training, is analyzed. The models are trained and tested

on a dataset collected over three months from a rooftop PV system in Sion, Switzerland, with a maximum

power of 22.2 kW. The Root Mean Square Error as well as the R

metrics are provided to assess the accuracy

of each method. The results demonstrate that both the choice of the actual input data and the sliding window

size signiﬁcantly inﬂuence the prediction accuracy. In particular, the results presented here show the potential

of combining different data sources to improve the accuracy of PV power prediction using LSTM models.

1 INTRODUCTION

1.1 Problem Deﬁnition

Solar power stands out due to its cleanliness, wide

availability, and scalability. However, photovoltaic

(PV) power output is inherently variable across mul-

tiple timescales due to factors such as geographical

location, as well as the size, orientation and technol-

ogy of the panels. Moreover, environmental condi-

tions like solar irradiance, temperature, cloud cover,

wind, and humidity crucially inﬂuence PV power out-

put (Pelland et al., 2013), (Pedro and Coimbra, 2012).

These characteristics pose signiﬁcant challenges for

grid stability and energy management (Chen and

Chang, 2021). Moreover, according to the World

Energy Outlook 2024 by the International Energy

Agency, global energy demand is expected to increase

by approximately 16% by 2050 under the Stated Poli-

cies Scenario (International Energy Agency, 2024).

However, if energy efﬁciency measures are not en-

https://orcid.org/0000-0001-6049-6536

https://orcid.org/0000-0003-3072-6075

https://orcid.org/0000-0002-9145-3157

hanced, this demand could rise by about 22% over

the same period. Accurate short-term PV power fore-

casts are thus critical not only for operational efﬁ-

ciency but also for effective energy storage, load con-

trol, and system reliability contributing to the broader

goals of energy conservation, sustainability, and cli-

mate change mitigation (Park et al., 2021). Conse-

quently, the development of robust predictive models

for PV power generation throughout the day is impor-

tant to address these challenges.

1.2 State of the Art

Prediction methods for PV power output span from

traditional approaches like persistence, physical, and

statistical models, which use both current PV data

and weather data, to more sophisticated models that

employ artiﬁcial intelligence (AI). Persistence models

commonly rely on present data for immediate predic-

tions (Wang et al., 2021), (Dash et al., 2021), while

physical models incorporate weather variables (Lima

et al., 2016). Statistical models (Wang et al., 2022),

such as ARMA (Benmouiza and Cheknane, 2016)

and ARIMA (Reikard, 2009), struggle with nonsta-

tionary and nonlinear data, limitations that more ad-

Jafari, F., Moerschell, J. and Riesen, K.

Predicting Photovoltaic Power Output Using LSTM: A Comparative Study Using both Historical and Climate Data.

DOI: 10.5220/0013258000003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 733-740

ISBN: 978-989-758-730-6; ISSN: 2184-4313

733

vanced AI models aim to overcome (Kumbhar et al.,

2021).

AI technologies, including machine learning

(ML), are important to improve building energy man-

agement systems as well as energy efﬁciency. ML

can broadly be divided into several branches, includ-

ing supervised and unsupervised learning, reinforce-

ment learning, and deep learning (Mellit et al., 2020).

Among these, supervised learning is the most preva-

lent approach in ML, where models are trained on la-

beled datasets that include both predictors and pre-

dictands (Markovics and Mayer, 2022). The over-

all goal of supervised learning is to identify the op-

timal functional relationship between input variables

and outputs, enabling accurate predictions. Super-

vised ML problems can generally be divided into two

categories, viz. classiﬁcation, which deals with cate-

gorical outputs, and regression, which addresses con-

tinuous outputs. PV power prediction, speciﬁcally, is

categorized as a regression problem due to its focus

on predicting continuous power output values (Wang

et al., 2023).

Common ML models used for PV power predic-

tions include artiﬁcial neural networks (ANN), sup-

port vector regression (SVR), k-nearest neighbors

(KNN) regression, as well as linear regression (Mel-

lit et al., 2020). While ANN models can capture

complex nonlinear relationships by utilizing large

datasets, they may often not fully address the dy-

namic characteristics of PV power generation, which

is inﬂuenced by various temporal factors (Meenal and

Selvakumar, 2018). As a result, alternative methods

such as SVR and KNN also attempt to model these

nonlinear relationships, but they can struggle with

the inherent time-dependent nature of the data (Wang

et al., 2017), (Liu et al., 2018), and (Mohammadi

et al., 2015). Deep learning models like recurrent

neural networks (RNNs) and long-short term memory

(LSTM) effectively address these challenges by cap-

turing sequential dependencies in data. While RNNs

excel in time-series prediction, they struggle with is-

sues like vanishing gradients, LSTMs may overcome

these problems, thereby enhancing the accuracy of

PV power prediction (Wang et al., 2020), (Wang et al.,

2019).

The superiority of LSTM models over other ML

methods for the prediction of PV power has been em-

pirically proven in several studies. In (Gao et al.,

2019), for instance, two distinct LSTM-based predic-

tion models tailored for ideal and non-ideal weather

conditions are introduced. For ideal weather, the

model leverages numerical weather prediction data

with seasonal adjustments, achieving an RMSE ac-

curacy of 4.62%. For non-ideal weather conditions,

the model integrates a discrete grey model to predict

daily total power, enhancing the LSTM’s accuracy in

scenarios such as rainy, cloudy, and overcast days. In

a second study, a one-hour ahead prediction of PV

power using historical PV power data only is pro-

posed (Abdel-Nasser and Mahmoud, 2019). The au-

thors compare the performance of the LSTM model

with three PV prediction methods for stationary time

series. Then, in (Hu et al., 2024) historical PV power

data and climate data is used separately to predict

the PV power. Results show that the use of climate

data leads in general to better results with a smaller

RMSE and the highest R

values. Finally, the impor-

tance of selecting predictors based on model archi-

tecture in solar energy forecasting is highlighted in

(Ciobanu et al., 2024). A dual-view LSTM model,

using both historical and future weather data, out-

performs a single-view model that only uses histor-

ical PV production and past weather data. Humidity

consistently emerges as an important predictor across

both models, underscoring the value of future weather

data in enhancing predictive accuracy.

1.3 Contribution

While several studies demonstrate the potential of

LSTMs in the domain of PV power prediction, they

often focus on single data sources or overlook the in-

ﬂuence of input data conﬁgurations on prediction ac-

curacy (e.g., the sliding window size). The present

paper provides a time-series prediction that seeks to

bridge these gaps. First, we thoroughly evaluate the

performance of LSTM models trained with differ-

ent data sources—past PV power, climate data, and

their combination—across multiple prediction hori-

zons and discuss on how these inputs effect the pre-

dictions. Second, the study also examines the effect

of varying sliding window sizes on the prediction ac-

curacy. That is, we provide a nuanced understanding

of how temporal context inﬂuences predictive perfor-

mance by providing empirical evidence using real-

world PV power data (adding practical relevance to

the ﬁndings of this paper).

1.4 Paper Organization

The remainder of this paper is organized as follows.

Section 2 outlines the methodology of the work, in-

cluding data type and preparation. Section 3 presents

the predictive model architecture. In Section 4 we

show and discuss the experimental results. Finally,

Section 5 concludes the paper with key insights.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

734

2 METHODOLOGY

2.1 Long-Short Term Memory Network

(LSTM)

One of the major advantages of RNNs is their ca-

pacity to incorporate contextual information when

mapping between input and output sequences, mak-

ing them particularly suitable for sequence predic-

tion tasks. However, traditional RNNs face a limi-

tation in effectively accessing long-term context, as

the inﬂuence of a given input decays or ampliﬁes ex-

ponentially with each recurrence, leading to what is

known as the vanishing or exploding gradient prob-

lem (Graves, 2012). Figure 1 shows a simple RNN

containing a single, self connected hidden layer.

Figure 1: A simple RNN containing a single, self connected

hidden layer (Graves, 2012).

LSTM networks (Hochreiter and Schmidhuber,

1997) are a type of RNN designed to handle long-

term dependencies using memory blocks instead of

traditional neurons. These blocks incorporate forget,

input, and output gates to manage the ﬂow of informa-

tion, addressing gradient decay and explosion issues.

This makes LSTMs highly effective for sequence pre-

diction tasks, such as PV power prediction, where

capturing long-term patterns is crucial.

Figure 2(a) illustrates a single-cell LSTM mem-

ory block, showing how an LSTM network structure

differs from a standard RNN by substituting the sum-

mation units in the hidden layer with memory blocks,

as seen in Figure 2(b). Although ordinary summa-

tion units can be combined with LSTM blocks, this

is generally unnecessary. Additionally, LSTM net-

works utilize the same output layers as those in stan-

dard RNNs (Graves, 2012).

Training of LSTM networks is conducted through

backpropagation through time, which allows the net-

work to learn from sequential data by minimizing the

prediction error over multiple time steps.

2.2 Description of the PV Power System

In this section we brieﬂy describe the PV power gen-

eration system used in this study, located at the au-

thors’ institution rooftop in Sion, Switzerland. Part of

the setup is shown in Figure 3. The PV system has an

installed maximum power of 22.2 kW and comprises

multiple solar panels positioned to capture maximum

sunlight throughout the day. The PV array consists of

60 individual solar panels, each with a rated power of

370 W. The PV modules use mono-crystalline silicon

technology. There are two AMPT converters feeding

a DC bus operating at 760 V. This DC bus connect

different loads and storage batteries. Table 1 summa-

rizes the speciﬁcations of an individual PV panel.

2.3 Data Collection and Preprocessing

This study uses real-time climate data from Me-

teoSwiss, including solar irradiance, temperature,

wind speed, humidity, and sunshine duration,

recorded at 10-minute intervals in Sion, Switzer-

land, from July 5 to September 30, 2024. The PV

power production data, unique to this research, is ob-

tained from a platform developed at GridLab, HES-

SO Valais, and synchronized with the climate data

for time-series analysis. Comprehensive data prepro-

cessing is manually conducted, including checks for

missing values, synchronization of climate data with

PV power output, and normalization. This study also

considers timestamps as features to train the model.

2.4 Sliding Window and Prediction

Horizons

Major goal of this paper is to predict the PV power

output at different prediction horizons, viz. 10, 30,

and 50 minutes into the future. These prediction

horizons are selected to provide short-term forecasts,

which are valuable for real-time grid management and

energy planning. To achieve these predictions, we uti-

lize a sliding window approach in the LSTM model,

where the length of historical data used for prediction

(i.e., the sliding window) varies between 60 and 20

minutes. The chosen window size potentially inﬂu-

ences the model’s accuracy, as it dictates the amount

of past information utilized to make each prediction.

Moreover, it is highly dependent on the type of data

actually employed to predict the PV power. In this

study, we consider three different types of data:

• Climate data

• Historical PV power data

• Combination of climate data and historical PV

power data

Predicting Photovoltaic Power Output Using LSTM: A Comparative Study Using both Historical and Climate Data

735

Figure 2: An LSTM network, (a) memory block with one cell, (b) with four input units, a hidden layer of two single-cell

LSTM memory blocks and ﬁve output units (Graves, 2012).

Table 1: Speciﬁcations of an individual PV panel used in this study.

Modules P

max

(W ) V

(V ) I

(A) V

mpp

(V ) I

mpp

(A)

LG370Q1C-V5 370 42.8 10.82 37.0 10.01

Figure 3: The studied system in Sion, Switzerland.

2.5 Evaluation Metrics

The Root Mean Square Error (RMSE) serves as a

widely recognized metric for regression model evalu-

ation. The RMSE is, for instance, employed as a stan-

dard statistical metric to assess model performance in

ﬁelds such as meteorology, air quality, and climate

research (Hodson, 2022). For a sample of n observa-

tions y (y

, i = 1, 2, ..., n), and n corresponding model

predictions ˆy, the RMSE is deﬁned as

RMSE =

∑

i=1

− ˆy

)

(1)

For PV power prediction, RMSE provides a clear

indication of the overall model accuracy (the smaller

the RMSE value, the better the model).

The coefﬁcient of determination R

is another

widely used metric in regression model evaluation

due to its intuitive interpretability (despite not always

being the primary criterion for model selection). For-

mally, the R

coefﬁcient is deﬁned as follows

= 1 −

∑

i=1

− ˆy

)

∑

i=1

− ¯y)

, (2)

where y

represents the observed values, ˆy

the pre-

dicted values, ¯y is the mean of the observed values,

and n is the number of observations.

The R

coefﬁcient indicates the proportion of vari-

ance in the observed data that is predictable from the

independent variables, making it valuable in under-

standing the model’s real-world applicability (Gao,

2024). An R

value closer to 1 suggests a model that

better captures the underlying patterns of the data.

3 PROPOSED WORK:

PREDICTIVE MODEL

This section describes the modeling approach for pre-

dicting PV power output using various input data

sources and the LSTM model’s structure. It outlines

three prediction models designed to assess the im-

pact of different data sources on prediction accuracy,

as discussed in Section 2.4. Additionally, it details

the LSTM model architecture and training procedures

speciﬁc to each model.

Figure 4 presents an overview of the entire mod-

eling process in a ﬂowchart format. The modeling

process is split in two major phases, the data prepro-

cessing and the PV power prediction. The PV power

prediction phase is thoroughly explained in the next

subsection.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

736

The preprocessing consists of two steps. The

ﬁrst step involves checking for missing entries in

both the PV power output and climate data to iden-

tify gaps and address inconsistencies. In the climate

dataset, recorded at 10-minute intervals, missing val-

ues are rare and only occur between two recorded

points. These missing entries are imputed with the

mean of the preceding and succeeding values, ensur-

ing a smooth temporal continuity in the climate data.

PV power output data, is originally recorded every

minute. Thus, the second step consist of synchroniza-

tion with the climate data to maintain consistent in-

tervals across datasets. To match the 10-minute inter-

vals of the climate data, the average PV power is cal-

culated over each 10-minute span, starting from the

same timestamp as the climate dataset.

Figure 4: Flowchart of the LSTM model.

3.1 Prediction Models

Three models are developed to predict PV power out-

put for 10, 30, and 50-minute horizons using different

input data to assess their impact on prediction accu-

racy.

3.1.1 Model 1: PV Power-Only Prediction

Model 1 uses historical PV power data as the sole

input for predicting future PV power output. By

leveraging the temporal patterns in past PV power

values, this model aims to understand the degree to

which prior PV output data alone can predict short-

term power generation. This model assumes that PV

power generation is somewhat self-correlated, captur-

ing daily cycles and other temporal patterns without

the inﬂuence of weather variables.

3.1.2 Model 2: Climate Data-Only Prediction

Model 2 is based exclusively on climate data, which

includes features such as solar irradiance, tempera-

ture, wind speed, humidity, and sunshine duration.

These features are collected from the MeteoSwiss

database at a 10-minute interval and synchronized

with the actual PV output data. By using climate

data only, this model evaluates the predictive power

of environmental conditions on PV generation, inde-

pendent of past PV power values. Climate data is of-

ten highly correlated with PV output, making it a key

predictor of solar energy production.

According to the correlation analysis provided in

Figure 5, we can see that irradiance and sunshine du-

ration are the most effective inputs in this endeavor

with correlations of about 0.95 and 0.80, respectively.

Figure 5: Correlation map between climate data and PV

power production.

3.1.3 Model 3: Combined Data Prediction

Model 3 combines both past PV power and climate

data to improve the prediction accuracy by integrat-

ing both temporal patterns and environmental inﬂu-

ences. This approach leverages the complementary

information between historical PV production data

and weather conditions, aiming to capture complex

relationships that could affect the PV output.

3.2 LSTM Model Architecture

Each of the three models shares a similar LSTM ar-

chitecture, differing only in the input layer based on

the data conﬁguration. The architecture includes four

LSTM layers with 50 hidden units, for learning tem-

poral dependencies and capturing patterns. A fully

connected layer translates LSTM features to the re-

gression output layer, which produces continuous PV

power predictions.

3.3 Training Procedure

Each model is trained using backpropagation through

time to optimize the weights, with the goal of mini-

mizing the prediction error. The complete dataset is

divided into training and testing subsets, with 90% of

the data used for training and 10% for testing. To

train the LSTM model, the Adam optimizer is used

(known for its efﬁciency and adaptability, especially

in handling noisy data and sparse gradients (Diederik,

2014)). The model is trained for a maximum of 200

epochs with an initial learning rate of 0.01 to balance

Predicting Photovoltaic Power Output Using LSTM: A Comparative Study Using both Historical and Climate Data

737

between convergence speed and stability during the

training.

4 RESULTS AND DISCUSSION

This section presents an analysis of the performance

of the three LSTM models under the discussed con-

ﬁgurations.

4.1 Estimation Model

The performance of Model 2, which predicts PV

power using only current climate data (horizon = 0), is

analyzed. Figure 6 shows estimated versus actual PV

power, with stem plots and RMSE values quantifying

performance. The model accurately distinguishes be-

tween day and night, with no errors at night, and fol-

lows daytime PV power closely, maintaining a mean

deviation below 1.5 kW in over 90% of cases. This

proof of concept demonstrates that estimating current

PV power using weather data is feasible. Future ex-

periments will explore prediction horizons of 10, 30,

and 50 minutes.

Figure 6: Results of the LSTM model to estimate the PV

power with climate data and time stamps tested for 9 days

with the stem plot and an RMSE of about 1kW.

4.2 Comparative Analysis of Prediction

Models

The effectiveness of each prediction model is evalu-

ated by comparing actual versus predicted PV power

values for models 1, 2, and 3, as shown in Figures

7, 8, and 9. All models use the past 20 minutes

of PV power data, climate data, their combination,

and timestamps to predict PV power for a 10-minute

horizon. While all models perform well, the model

combining historical data with climate data generally

yields the best results in qualitative analysis.

The three LSTM models are also used to predict

PV power for 30 and 50 minutes ahead. The RMSE

and R

values are shown in Figure 10 and 11, re-

spectively. The results indicate that, across all pre-

diction horizons, utilizing a combination of histori-

cal PV power data and climate data consistently leads

to more accurate PV power predictions. That is, this

combined approach appears to enhance the model’s

ability to capture both short-term trends from recent

PV output and underlying patterns inﬂuenced by cli-

matic conditions, resulting in generally improved pre-

diction accuracy.

Figure 7: A LSTM model to predict the PV power with

historical data and timestamps with prediction horizon of

10 minutes.

Figure 8: A LSTM model to predict the PV power with

climate data and timestamps with prediction horizon of 10

minutes.

Figure 9: A LSTM model to predict the PV power with his-

torical, climate data and timestamps with prediction horizon

of 10 minutes.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

738

Table 2: Performance metrics for PV power prediction using a combination of historical and climate data across two sliding

window sizes are presented, with the best results highlighted in bold.

Combination of both data

Window sizes

Predict 10 min ahead Predict 30 min ahead Predict 50 min ahead

RMSE (%) R

20 min 4.39 0.93 6.25 0.86 6.67 0.84

60 min 4.50 0.93 6.52 0.85 7.19 0.82

Figure 10: RMSE values for different LSTM models to pre-

dict the PV power.

Figure 11: R

values for different LSTM models to predict

the PV power, starting from 0.6.

4.3 Impact of Sliding Window Size

An additional experiment evaluates the impact of a

60 minutes sliding window size compared to the 20-

minute window used previously. The results of this

experiment are shown in Table 2, where we present

the RMSE and R

values for both window sizes,

namely 20 and 60 minutes (for the sake of simplic-

ity, we only show the results of combining both in-

puts, historical data, and climate data – similar effects

are to be expected for the individual data sets). We

ﬁnd that for small prediction horizon (20-minute), the

results are slightly better. However, for longer predic-

tion horizon (60-minute), the higher RMSE and lower

values for the 60-minute window size are clear.

The nonlinear and highly oscillatory nature of climate

data apparently results in suboptimal model training

at long prediction horizons when a larger moving win-

dow is used. Perhaps when a shadow passes, it is

more important to detect the start of the shadow than

to remember what happened in the last 60 minutes

(and the beginning of the shadow can best be recog-

nized based on the current and past PV power gen-

eration values). Overall, these results suggest that

a smaller sliding window can better capture relevant

short-term dependencies when predicting PV power

in longer prediction horizons.

5 CONCLUSIONS

This paper investigates the accuracy of PV power

prediction using LSTM models trained with differ-

ent datasets, namely historical PV power data, climate

data, a combination of both all with timestamps. Pre-

diction horizons of 10, 30, and 50 minutes are ex-

plored to assess model performance under varying

conditions, and two different sliding window sizes

are considered to show the effect of time and data

type in prediction. Results indicate that RMSE val-

ues increase as the prediction horizon extends. The

combined dataset of PV power and climate data with

timestamps consistently yields lower RMSE values

compared to using either dataset individually, indicat-

ing that integrating diverse data sources enhances pre-

diction accuracy. The results also underscore the im-

portance of selecting an appropriate sliding window

size. These ﬁndings contribute valuable insights into

optimizing LSTM models for short-term PV power

prediction, with practical implications for real-world

solar energy prediction.

ACKNOWLEDGEMENTS

This research is funded by the Swiss National Science

Foundation (SNSF) under grant no. 10CE-1

216893

in cooperation with CETPartnership, under Joint Call

2022 for research proposals, Increasing control and

efﬁciency in regional energy systems using quantum

sensors and machine learning – QuantumIRES, UE-

FISCDI PNCDI IV.

Predicting Photovoltaic Power Output Using LSTM: A Comparative Study Using both Historical and Climate Data

739

REFERENCES

Abdel-Nasser, M. and Mahmoud, K. (2019). Accurate pho-

tovoltaic power forecasting models using deep LSTM-

RNN. Neural Comput. Appl., 31(7):2727–2740.

Benmouiza, K. and Cheknane, A. (2016). Small-scale so-

lar radiation forecasting using arma and nonlinear au-

toregressive neural network models. Theoretical and

Applied Climatology, 124:945–958.

Chen, H. and Chang, X. (2021). Photovoltaic power predic-

tion of lstm model based on pearson feature selection.

Energy Reports, 7:1047–1054.

Ciobanu, A., Bacanin, N., Sanders, T., and Stoean, C.

(2024). Assessing predictor inﬂuence in lstm mod-

els for enhanced solar energy forecasting. In Pro-

ceedings of the 26th International Symposium on Sym-

bolic and Numeric Algorithms for Scientiﬁc Comput-

ing (SYNASC).

Dash, D. R., Dash, P., and Bisoi, R. (2021). Short term solar

power forecasting using hybrid minimum variance ex-

panded rvﬂn and sine-cosine levy ﬂight pso algorithm.

Renewable Energy, 174:513–537.

Diederik, P. K. (2014). Adam: A method for stochastic

optimization. (No Title).

Gao, J. (2024). R-squared (r2)–how much variation is ex-

plained? Research Methods in Medicine & Health

Sciences, 5(4):104–109.

Gao, M., Li, J., Hong, F., and Long, D. (2019). Day-ahead

power forecasting in a large-scale photovoltaic plant

based on weather classiﬁcation using lstm. Energy,

187:115838.

Graves, A. (2012). Long short-term memory. Supervised

sequence labelling with recurrent neural networks,

pages 37–45.

Hochreiter, S. and Schmidhuber, J. (1997). Long short-term

memory. Neural Comput., 9(8):1735–1780.

Hodson, T. O. (2022). Root mean square error (rmse)

or mean absolute error (mae): When to use them or

not. Geoscientiﬁc Model Development Discussions,

2022:1–10.

Hu, Z., Gao, Y., Ji, S., Mae, M., and Imaizumi, T. (2024).

Improved multistep ahead photovoltaic power pre-

diction model based on lstm and self-attention with

weather forecast data. Applied Energy, 359:122709.

International Energy Agency (2024). World energy outlook

2024.

Kumbhar, A., Dhawale, P. G., Kumbhar, S., Patil, U., and

Magdum, P. (2021). A comprehensive review: Ma-

chine learning and its application in integrated power

system. Energy Reports, 7:5467–5474.

Lima, F. J., Martins, F. R., Pereira, E. B., Lorenz, E., and

Heinemann, D. (2016). Forecast for surface solar irra-

diance at the brazilian northeastern region using nwp

model and artiﬁcial neural networks. Renewable En-

ergy, 87:807–818.

Liu, L., Zhao, Y., Chang, D., Xie, J., Ma, Z., Sun, Q., Yin,

H., and Wennersten, R. (2018). Prediction of short-

term pv power output and uncertainty analysis. Ap-

plied energy, 228:700–711.

Markovics, D. and Mayer, M. J. (2022). Compari-

son of machine learning methods for photovoltaic

power forecasting based on numerical weather pre-

diction. Renewable and Sustainable Energy Reviews,

161:112364.

Meenal, R. and Selvakumar, A. I. (2018). Assessment of

svm, empirical and ann based solar radiation predic-

tion models with most inﬂuencing input parameters.

Renewable Energy, 121:324–343.

Mellit, A., Massi Pavan, A., Ogliari, E., Leva, S., and Lughi,

V. (2020). Advanced methods for photovoltaic out-

put power forecasting: A review. Applied Sciences,

10(2):487.

Mohammadi, K., Shamshirband, S., Anisi, M. H., Alam,

K. A., and Petkovi

c, D. (2015). Support vector re-

gression based prediction of global solar radiation on

a horizontal surface. Energy Conversion and Manage-

ment, 91:433–441.

Park, M. K., Lee, J. M., Kang, W. H., Choi, J. M., and Lee,

K. H. (2021). Predictive model for pv power genera-

tion using rnn (lstm). Journal of Mechanical Science

and Technology, 35(2):795–803.

Pedro, H. T. and Coimbra, C. F. (2012). Assessment of fore-

casting techniques for solar power production with no

exogenous inputs. Solar Energy, 86(7):2017–2028.

Pelland, S., Remund, J., Kleissl, J., Oozeki, T., and De Bra-

bandere, K. (2013). Photovoltaic and solar forecast-

ing: state of the art. IEA PVPS Task, 14(355):1–36.

Reikard, G. (2009). Predicting solar radiation at high reso-

lutions: A comparison of time series forecasts. Solar

energy, 83(3):342–349.

Wang, F., Lu, X., Mei, S., Su, Y., Zhen, Z., Zou, Z., Zhang,

X., Yin, R., Dui

c, N., Shaﬁe-khah, M., et al. (2022).

A satellite image data based ultra-short-term solar pv

power forecasting method considering cloud informa-

tion from neighboring plant. Energy, 238:121946.

Wang, F., Xuan, Z., Zhen, Z., Li, K., Wang, T., and

Shi, M. (2020). A day-ahead pv power forecasting

method based on lstm-rnn model and time correla-

tion modiﬁcation under partial daily pattern predic-

tion framework. Energy Conversion and Manage-

ment, 212:112766.

Wang, F., Zhen, Z., Wang, B., and Mi, Z. (2017). Compar-

ative study on knn and svm based weather classiﬁca-

tion models for day ahead short term solar pv power

forecasting. Applied Sciences, 8(1):28.

Wang, H., Lei, Z., Zhang, X., Zhou, B., and Peng, J.

(2019). A review of deep learning for renewable en-

ergy forecasting. Energy Conversion and Manage-

ment, 198:111799.

Wang, L., Mao, M., Xie, J., Liao, Z., Zhang, H., and Li,

H. (2023). Accurate solar pv power prediction in-

terval method based on frequency-domain decompo-

sition and lstm model. Energy, 262:125592.

Wang, M., Peng, J., Luo, Y., Shen, Z., and Yang, H. (2021).

Comparison of different simplistic prediction models

for forecasting pv power output: assessment with ex-

perimental measurements. Energy, 224:120162.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

740