DEVELOPMENT OF AN ENHANCED ALBIRUNI OPTIMIZED LONG SHORT TERM

MEMORY (LSTM) NETWORK FOR ACCURATE PREDICTION OF RAINFALL IN

NIGERIAS

[Link]. RESEARCH PROPOSAL

BY

Salisu Zubairu
P17PSCS8016

Supervisors
Prof. Abdulkarim Muhammad
Prof. Abdullahi Muhammad

i
 DECLARATION

I Salisu Zubairu, hereby declare that this [Link]. proposal titled “Development of an Enhanced

Albiruni Optimized LSTM Network for Accurate Prediction of Rainfall in Nigeria” is being

carried out by me under the supervision of Prof. Abdulkarim Muhammad and Prof. Abdullahi

Muhammad. This research work has not been presented for the award of any degree in any

institution before now. All borrowed information have been duly acknowledged.

___________________ _________________ _______________

Salisu Zubairu Signature Date

ii
 CERTIFICATION
This [Link]. proposal titled “Development of an Enhanced Albiruni Optimized LSTM Network

for Accurate Prediction of Rainfall in Nigeria” by Salisu Zubairu P17PSCS8016 meets the

regulations governing the award of the degree of [Link]. in Computer Science of the Ahmadu

Bello University, Zaria, and is approved for its contribution to knowledge and literary

presentation.

Prof. Abdulkarim Muhammad ________________________

Major, Supervisory Committee Signature and date

Prof. Abdullahi Muhammad ________________________

Minor, Supervisory Committee Signature and date

Prof. A.F.D Kana ________ ________________

Head of Department Signature and date

iii
 Table of Content
Table of Content.............................................................................................................................iv

List of Figures................................................................................Error! Bookmark not defined.

Abstract...........................................................................................................................................iv

Introduction......................................................................................................................................1

Problem Statement...........................................................................................................................2

Objectives........................................................................................................................................3

Aim..................................................................................................................................................3

Literature Review............................................................................................................................4

Methodology ……………………………………………………………………………………...8

Research Organization.....................................................................................................................8

References......................................................................................................................................10

iv
 ABSTRACT

Rainfall prediction remains a complex yet crucial component of environmental forecasting,

particularly in countries that heavily depend on rain-fed agriculture and are frequently affected
by extreme weather events. Accurate and timely rainfall prediction c
an significantly enhance agricultural productivity, inform water resource management strategies,
and strengthen disaster preparedness. The reliance on traditional methods for forecasting in many
developing nations often leads to delayed responses, poor planning, and increased vulnerability
to climate-induced disasters such as floods due to the limitations of the models to capture
complex patterns in time series datasets underscoring the urgency for more reliable predictive
tools. Recently, researchers have developed metaheuristics algorithms such as Albiruni Earth
Radius Algorithm (AERA) to improve the performance of the long short term memory (LSTM),
a variant of recurrent neural networks known for its efficacy in modeling time-series data.
However, the algorithm suffers entrapment in the local optima. This research will integrate
Albiruni Earth Radius Algorithm (AERA) and Levy Flight Algorithm (LFA) to allow the LSTM
effectively navigate the solution space and converge to optimal parameters for improved
prediction accuracy. To establish a comparative baseline, the research utilized an LSTM model
optimized using the Albiruni Earth Radius Algorithm for rainfall prediction in Ethiopia using
dataset for Jimma region. The performance of the enhanced model will be evaluated using
standard regression metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and the
Coefficient of Determination (R²).

v
 CHAPTER ONE
INTRODUCTION

1.0 Background to the Study

Artificial intelligence (AI) has emerged as a transformative technology with widespread

applications across various domains. One of its key advancements is the development of

Artificial Neural Networks (ANNs), which enable machines to process and analyze complex data

by mimicking human cognitive functions. A specialized form of ANNs, Recurrent Neural

Networks (RNNs), excel in handling sequential data, making them particularly useful for tasks

such as time-series prediction. In such applications, RNNs leverage labelled datasets to generate

forecasts, supporting decision-making in fields like meteorology. For instance, experts can

employ these models to predict rainfall patterns, which are critical for agricultural planning,

disaster preparedness, and water resource management. Among the various deep learning

architectures used for sequential data analysis, Long Short-Term Memory (LSTM) networks a

subset of machine learning within the broader AI framework have proven highly effective. Their

ability to capture long-term dependencies in time-series data makes them well-suited for rainfall

prediction, where patterns may exhibit subtle variations over time. Traditional methods of

analyzing meteorological data often struggle to detect these minor fluctuations, even when

interpreted by expert meteorologists. However, AI-driven approaches, particularly those

incorporating deep learning techniques, have significantly improved the accuracy and efficiency

of rainfall forecasting. Recent advancements in deep learning have facilitated the development of

sophisticated models capable of identifying hidden patterns in historical climate data. These

innovations empower meteorologists to make more informed and timely decisions regarding

rainfall events. Despite these improvements, there remains a need for further optimization to

enhance predictive performance, particularly in region-specific contexts. In this study, an

enhanced Al-Biruni-optimized Long Short-Term Memory (LSTM) network is proposed for

vi
rainfall prediction in Nigeria. By integrating metaheuristic optimization techniques with deep

learning, this model aims to improve forecasting precision, thereby supporting better climate-

related decision-making in Nigeria. Recently, researchers have focused on the optimization of

LSTM parameters for the rainfall prediction.

Some metaheuristic optimization that had been employed in this direction include Particle swarm

optimization (PSO) algorithm, and the Albiruni optimization algorithm (El-Sayed, et al., 2023)

among others. Although the root mean squared error (RMSE) obtained by the Albiruni optimized

LSTM is lower compared to that obtained by other LSTM based algorithms it was compared

with due to local optima problem, the error could still be further decreased.

1.1 Problem Statement

Rainfall prediction is a vital task for various sectors in Nigeria, particularly agriculture, water resource

management, and disaster preparedness. As the country experiences diverse climatic conditions, ranging

from arid regions in the north to tropical rainforests in the south, accurately forecasting rainfall patterns is

critical for mitigating the risks of floods, droughts, and food insecurity. Traditionally, rainfall forecasting

has relied on physical models and statistical techniques, but these methods often struggle with the chaotic

and non-linear nature of weather data. Recently, machine learning models, especially Long Short-Term

Memory (LSTM) networks, have emerged as a leading approach for predicting rainfall due to their ability

to capture temporal dependencies and non-linear patterns in time-series data. To further improve the

predictive performance of LSTM, several metaheuristic algorithms had been employed to tune its

weights, one of these algorithms is the Albiruni algorithm (El-Sayed, et al., 2023). However, despite the

improvement in predictive performance obtained using this algorithm, the Root Mean Square Error

(RMSE) obtained is still high due to the fact that, the Albiruni algorithm suffers from local optima

problem (Bofan, et al., 2023).

This research will enhance Albiruni LSTM thereby adding to existing methods for accurate

prediction of rainfall.

vii
1.2 AIM AND OBJECTIVES

Aim

The aim of this research is to enhance and implement a long short term memory network that

will effectively predict rainfall with the highest possible accuracy while limiting time

complexity.

Objectives

The specific objectives of the research are:

a. To implement the LSTM algorithm and apply the enhanced Albiruni algorithm on it.

b. To enhance the Albiruni algorithm by addressing its local optima problem using levy

flight algorithm.

c. To evaluate the enhanced Albiruni optimized LSTM on the prediction of rainfall and

compare the result obtained with the work of (El-Sayed, et al., 2023) and other LSTM

variants.

viii
 CHAPTER TWO

Literature Review
2.1 Introduction

Rainfall prediction is a crucial aspect of meteorology, impacting agriculture, water resource

management, disaster mitigation, and climate change adaptation. Traditional forecasting methods

have been replaced or augmented by machine learning techniques, with Long Short-Term

Memory (LSTM) networks emerging as one of the most effective approaches due to their ability

to handle sequential data. Recent advancements in optimization techniques, such as the Al-Biruni

Optimization Algorithm (ABOA), have further improved predictive accuracy.

2.2 The LSTM Network

LSTM has been found by many researchers in the literature to be good for processing time series

sequences especially if these sequences have long term dependencies. An LSTM unit has an

input gate (I), a forget gate (F) and an output gate (O). They regulate the transmission of

information in the network. LSTM can be modelled mathematically as shown below. The output

of the forget gate F n.

F n=σ ( W F . ( h t−1 . x t ) ). b I equation 1

where, 𝑊. and 𝑏. indicate the weight and bias, respectively, for the corresponding gate.

I t=σ ( W I . ( ht−1 . x t ) ). b F

~
C t =tanh ( W c . ( ht −1 . x t ) ). b c

C t = F t .C t −1 + I t.~
Ct

ix
2.3 Review of Related Work

Long Short-Term Memory (LSTM) networks, a type of Recurrent Neural Network (RNN), have

consistently demonstrated superior performance in time series forecasting tasks. Introduced by

(Hochreiter & Schmidhuber 1997), LSTM was proposed to address the vanishing gradient

problem in standard RNNs, enabling better learning of long-term dependencies. Subsequent

studies have validated the use of LSTM in various domains of time series prediction.

For instance, (Kratzert et al., 2018) applied LSTM in hydrological modelling and found that it

significantly outperformed traditional models such as ARIMA and decision trees. (Zaytar &

Amrani 2016) also employed LSTM for rainfall prediction and observed better performance in

capturing long-term dependencies compared to classical models. (Shahid et al. 2022)

demonstrated that deep learning models, particularly LSTM, consistently provided higher

accuracy in rainfall prediction tasks than shallow learning methods like support vector machines

(SVM) and random forests. Their findings emphasize the value of LSTM in learning complex

temporal patterns. Similarly, (Tian & Pan 2020) explored multivariate LSTM models and

reported superior prediction accuracy in short-term meteorological forecasting. Hyper parameter

tuning is another essential factor influencing LSTM performance. (Mustain et al., 2022)

reported that varying LSTM parameters such as hidden layers, batch size, and learning rate drop

periods can significantly enhance predictive performance, especially when evaluated using

metrics like the Mean Arctangent Absolute Percentage Error (MAAPE). Moreover, integrating

feature selection techniques like Principal Component Analysis (PCA) and stacking multiple

LSTM layers has led to improved forecasting accuracy (Zhao et al., 2021).

2.4 Metaheuristic Optimization Algorithm Based LSTM for Rainfall Prediction

Optimizing LSTM networks is crucial for improving their predictive accuracy and computational

efficiency. Several metaheuristic algorithms have been employed to tune LSTM hyper
x
parameters effectively. Techniques such as Genetic Algorithms (GA), Particle Swarm

Optimization (PSO), Grey Wolf Optimization (GWO), and Bayesian Optimization have been

widely explored for this purpose. (Fehaid, 2024) utilized a grid search-based LSTM model to

predict rainfall in Mecca, Saudi Arabia, achieving moderate accuracy but facing limitations due

to the fixed nature of grid search. In contrast, (El-Sayed, et al., 2023) proposed an Al-Biruni

Earth Radius Optimization (BER) algorithm for LSTM tuning, which outperformed traditional

LSTM configurations in solving regression problems related to rainfall prediction. However,

despite achieving lower Mean Squared Error (MSE), challenges remained due to the tendency of

certain algorithms to converge to local optima. Additional studies have explored hybrid

approaches.

For instance, (Khosravi et. al. 2020) combined PSO and GWO for LSTM optimization, resulting

in faster convergence and enhanced accuracy. Likewise, (Li. et al. 2021) demonstrated that using

metaheuristics not only improved LSTM's generalization but also reduced over fitting,

particularly in datasets with high seasonal variance. These findings underscore the growing

relevance of combining LSTM with intelligent optimization algorithms to handle the complex,

non-linear patterns in meteorological time series data effectively.

Standard Albiruni Algorithm

Al-biruni Optimization Algorithm by (El-Sayed, et al., 2023)

1. Initialize population ⃗
Si ( i=1 ,2 … … . , d ) with size d. Maximum iteration Max iter
,Fitness function F n
2. Initialize the algorithm’s parameters.
3. Set t ¿ 1
4. Calculate Fitness function F n for each ⃗
Si

5. Find best solution ⃗

S¿
6. While t ≤ Max iter do

xi
7. FOR each solution in the exploration group do
8. Heading towards the best solution
cosc( x )
9. r =h
1−cosc (x )
10. D ¿ r⃗1 ( ⃗S ( t )−1 )
⃗

11. ⃗S ( t+1 ) ¿ ⃗S ( t )+ ⃗
D (2 ⃗
r 1−1 )
12. END FOR
13. FOR each solution in the exploitation group, do
14. Elitism of the best solution
15. ⃗ r2 ( ⃗
D ¿ ⃗ L ( t )−⃗S (t) )

16. ⃗
S1 (t +1)=r 2 ( S ( t ) + ⃗
D)
17. Investigating the area around the best solution
2
2∗t
18. k⃗ =1+ 2
Max iter

19. S2 ( t+1 ) =r ( ⃗
⃗ S¿ ( t ) + ⃗k )

20. Compare ⃗ S2 ( t +1 ) and select the best solution ⃗

S1 ( t +1 ) and ⃗ S¿
21. If best fitness didn’t change from previous two iterations, then
22. Mutate the solution
cosc (x)
23. ⃗S ( t+1 )=⃗k * z 2- h
1−cosc( x )
24. end if
25. END FOR
26. Update fitness F n for each ⃗S
27. END WHILE
28. Return best solution ⃗
S¿

Fig. 2.1

xii
 CHAPTER THREE

Research Methodology

3. 0 Methodology
Step I: Dataset Pre-processing

Historical meteorological data for Nigeria will be sourced from global repositories such as

Kaggle, the World Meteorological Organization (WMO), and satellite-based remote sensing

platforms. Key features will include daily/monthly rainfall, temperature, humidity, wind speed,

and atmospheric pressure.

• Data cleaning will involve the removal of entries with excessive missing values to reduce

noise and enhance data integrity.

• Feature normalization techniques may be applied to standardize the dataset.

The dataset will then be partitioned into training (70%) and testing (30%) sets to ensure robust

model validation.

Step II: Enhancement Using Levy Flight Random Walk

The standard Al-Biruni algorithm will be modified by integrating Levy flight random walk to

enhance exploration capabilities, prevent premature convergence, and improving optimization

during training.

⃗S ( t+1 )=⃗k ∗z 2−h cosc ( x ) +levyflight (d ) equation 2

1−cosc ( x )

Step III: Model Implementation

Both the optimized LSTM model and the enhanced Al-Biruni algorithm will be developed using

Py Torch, Mindspore and Tensor Flow frameworks.

xiii
Hyper parameter tuning and early stopping will be incorporated to optimize model performance

and prevent over fitting.

Step IV: Model Training and Validation

Models will be trained on the pre-processed dataset, and validation will be performed using the

test set. Cross-validation techniques may also be employed to ensure generalizability.

Step V: Performance Evaluation

Performance will be assessed using standard evaluation metrics: Nash–Sutcliffe Efficiency

(NSE), Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error

(MAE), and the coefficient of determination (R²), enabling comprehensive comparison across

models.

Conclusion
The application of LSTM networks for rainfall prediction has demonstrated significant

improvements over traditional methods. Optimization techniques, particularly metaheuristic

algorithms like ABOA had provide a viable solution for fine-tuning LSTM models.

xiv
 REFERENCES

Adefolalu, D. O. (2007). Climate change and economic sustainability in Nigeria. Journal of

Environmental Science, 12 (3), 45-59.

Ata, A. A., Tiantian, Y., Kuolin, H., Soroosh, S., Junqiang, L., & Qidong, P. (2018). Short-Term
Precipitation Forecast Based on the PERSIANN System and LSTM Recurrent Neural
Networks. Journal of Geophysical Research: Atmospheres, 1254-12563.
Bofan, H., Nurlida, I., Kimberley, K. K., & Gang, C. (2023). Techno-economic analysis of an
HRES with fuel cells, solar panels,and wind turbines using an improved Al-Biruni
algorithm. Heliyon 9 , 1-23.
El-Sayed, E.-k., Abdelaziz, A., Fadwa, A., Mostafa, A., Abdelhameed, I., & Doaa, S. K. (2023).
Al-Biruni Based Optimization of Rainfall Forecasting in Ethiopia. Computer Systems
Science & Engineering, 2885-2899.
Fehaid, A. (2024). AI-driven improvement of monthly average rainfall forecasting in Mecca
using grid search optimization for LSTM networks. Journal of Water and Climate
Change Vol 15 No 4, 1439 , 1439-1458. Li.
Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8),
1735-17-80.
Kratzert, F., Klotz, D., Shalev, G., et al. (2018). Rainfall-runoff modeling using LSTM networks.
Hydrology and Earth System Sciences, 22(11), 6005-6022.

Muhammad, A. F., Jasruddin, Pariabti, P., Helmi, Muhammad, A., & Hari, T. (2023). Monthly
Rainfall Prediction Using the Backpropagation Neural Network (BPNN) Algorithm in
Maros Regency. Scientific Journal of Informatics , 13-24.
Prasetya, E. P., & Esmeralda, C. D. (2019). Rainfall forecasting for the natural disasters
preparation using recurrent neural networks. International Conference on Electrical
Engineering and Informatics (pp. 52-57). IEEE.
Rohit, D. (2012). On The Rainfall Time Series Prediction Using Multilayer Perceptron Artificial
Neural Network . International Journal of Emerging Technology and Advanced
Engineering,Volume 2, Issue 1, 148-15.
Shahid, F., Zameer, A., & Muneeb, M. (2020). Predictions for COVID-19 using deep learning
models. Chaos, Solitons & Fractals, 140, 110212.
Zaytar, M. A., & Amrani, C. E. (2016). Sequence-to-sequence LSTM model for time series
prediction. International Conference on Big Data.

xv
Research Organization

The research is Organized as follows:

Chapter 1 is the introduction which presents the research background, problem statement, aim, objective
and significance of the research.

Chapter 2 is a literature review that describes the concepts and definition of the previous related studies.

Chapter 3 is a research methodology which covers the detailed description of the proposed research
design, approach and datasets description.

Chapter 4 discussion and analysis of results.

Chapter 5 discussion of findings, conclusion and recommendation.

xvi